CN110008575B - Multi-parameter predictive control algorithm for switching process medium multi-temperature target set values of circulating cooling water system - Google Patents

本发明公开了一种循环冷却水系统工艺介质多温度目标设定值切换多参数预测控制算法，具体步骤为：获取循环冷却水系统的历史数据；从历史数据中筛选出特征变量，经数据处理后分为训练样本集和测试样本集；对训练样本集进行逐层贪婪无监督预训练，得到深层神经网络初始化的输入层与隐含层的权值矩阵W、输入层与隐含层阈值矩阵B；进行参数微调，得到初始工艺介质温度预测模型；再测试、评估后；通过获取当前数据，得到冷却给回水压差设定值和工艺介质温度预测值。有益效果：精确降温，在高精度的同时还能最大限度节能。

The invention discloses a multi-temperature target setting value switching multi-parameter predictive control algorithm for a process medium in a circulating cooling water system. The specific steps are: acquiring historical data of the circulating cooling water system; After that, it is divided into training sample set and test sample set; the training sample set is subjected to layer-by-layer greedy unsupervised pre-training, and the weight matrix W of the input layer and the hidden layer initialized by the deep neural network, the threshold matrix of the input layer and the hidden layer are obtained B. Perform parameter fine-tuning to obtain the initial process medium temperature prediction model; after retesting and evaluation; obtain the current data to obtain the set value of the cooling water supply and return water pressure difference and the process medium temperature prediction value. Beneficial effects: accurate cooling, high precision and maximum energy saving.

循环冷却水系统工艺介质多温度目标设定值切换多参数预测 控制算法Multi-parameter Prediction of Multi-Temperature Target Setting Value Switching of Process Medium in Circulating Cooling Water System control algorithm

技术领域technical field

本发明涉及工业生产中循环冷却水技术领域，具体的说是一种循环冷却水系统工艺介质多温度目标设定值切换多参数预测控制算法。The invention relates to the technical field of circulating cooling water in industrial production, in particular to a multi-parameter predictive control algorithm for switching between multi-temperature target setting values of a process medium in a circulating cooling water system.

背景技术Background technique

在化工、电力、冶金等工业生产过程中，系统往往会因为摩擦、燃烧、化学反应等产生大量的热量，导致设备及系统温度升高，影响生产区工艺介质品质，造成严重的经济损失。循环冷却水系统是一种常见的在工业生产现场对工艺介质控温工程系统，利用传热介质将生产过程中产生的热量传导到自然环境中，达到降温的目的，应用广泛。In the industrial production process of chemical industry, electric power, metallurgy, etc., the system often generates a lot of heat due to friction, combustion, chemical reaction, etc., which will cause the temperature of equipment and system to rise, affect the quality of the process medium in the production area, and cause serious economic losses. The circulating cooling water system is a common engineering system for controlling the temperature of the process medium at the industrial production site. It uses the heat transfer medium to transfer the heat generated in the production process to the natural environment to achieve the purpose of cooling. It is widely used.

循环冷却水系统管网拓扑结构规模庞大，结构复杂，且其内部组件繁多，系统设计大多是依靠经验，为满足生产需要，通常依据最大负荷并给予一定富余量而盲目提高供水能力和冷却能力，与实际工业生产要求之间常存在“大马拉小车”现象，造成了大量的冷却资源浪费。因此，提出建立循环冷却水工艺介质温度节能控制系统，根据工业生产现场各换热器内工艺介质实时检测温度，确定换热现场实际需求的最小冷却水流量，减少循环冷却水系统能耗。由于冷却给回水压差与冷却水流量成正比，冷却给回水压差越大，意味着冷却水流量越大，而工业生产现场管网管径均较大(直径约2-3米)不便进行冷却水流量检测，因此用冷却给回水压差等效替代冷却水流量检测。The pipe network topology of the circulating cooling water system is large in scale, complex in structure, and has many internal components. Most of the system design is based on experience. In order to meet the production needs, the water supply capacity and cooling capacity are usually blindly increased based on the maximum load and a certain margin. There is often a phenomenon of "big horses and small carts" between actual industrial production requirements, resulting in a lot of waste of cooling resources. Therefore, it is proposed to establish an energy-saving control system for the process medium temperature of circulating cooling water. According to the real-time temperature detection of the process medium in each heat exchanger in the industrial production site, the minimum cooling water flow rate actually required by the heat exchange site is determined to reduce the energy consumption of the circulating cooling water system. Since the cooling water supply and return water pressure difference is proportional to the cooling water flow rate, the larger the cooling water supply and return water pressure difference means the greater the cooling water flow rate, and the pipe network diameter of the industrial production site is relatively large (about 2-3 meters in diameter) It is inconvenient to detect the cooling water flow, so the pressure difference between the cooling supply and return water is equivalent to replace the cooling water flow detection.

循环冷却水工艺介质温度节能控制系统具有大滞后、大惯性、非线性等特点。由于滞后的存在严重影响了系统的稳定性和控制性能，而基于精确数学模型的常规控制方法通常难以获得满意的动、静态控制性能，并且系统在运行中参数的时变和外界环境的不确定因素的影响下，使这种温度预测控制系统更加难以控制。The energy-saving control system for the temperature of the circulating cooling water process medium has the characteristics of large lag, large inertia, and nonlinearity. Because the existence of hysteresis seriously affects the stability and control performance of the system, conventional control methods based on precise mathematical models are usually difficult to obtain satisfactory dynamic and static control performance, and the time-varying parameters of the system and the uncertainty of the external environment during operation Under the influence of factors, it is more difficult to control the temperature predictive control system.

19世纪中期至20世纪50年代末称为“经典控制”时代，工业中使用较多的控制方式是以PID控制，或者和以经典理论为基础的反馈、前馈相组合的方式，但是对于具有大时滞、非线性、时变的系统控制中，基于传统的控制方法就很难保障控制系统的稳定性和控制精度。20世纪50中期至70年代称为“现代控制”时代，Smith预估控制算法是较早应用到时滞控制系统中的一种算法，到上个世纪90年代仍然有很多人应用各种改进的Smith控制算法。在国际上，时滞工业控制中主要采用的是最优控制、系统辨识、自适应控制等技术。如美国在六七十年代曾在轧机控制中引入了在线参数估计和离散化模型，并应用广义最小方差控制算法建立了冶金加热炉的自适应控制系统。20世纪70年代至今为“智能控制阶段”，各种智能控制理论如模糊控制、神经控制、专家控制等以及它们之间或者和传统控制理论的结合也已经广泛应用在温度智能控制中，提高了系统的自动控制能力，增强了系统温度性和鲁棒性。From the mid-19th century to the end of the 1950s, it was called the era of "classical control". The control method used more in industry was PID control, or combined with feedback and feedforward based on classical theory, but for those with In the system control of large time-delay, nonlinear and time-varying, it is difficult to guarantee the stability and control accuracy of the control system based on the traditional control method. From the mid-1950s to the 1970s, it was called the era of "modern control". The Smith predictive control algorithm was an algorithm that was applied to the time-delay control system earlier. In the 1990s, many people still applied various improved methods. Smith control algorithm. Internationally, technologies such as optimal control, system identification, and adaptive control are mainly used in time-delay industrial control. For example, the United States introduced online parameter estimation and discretization models in rolling mill control in the 1960s and 1970s, and applied the generalized minimum variance control algorithm to establish an adaptive control system for metallurgical heating furnaces. From the 1970s to the present is the "intelligent control stage". Various intelligent control theories such as fuzzy control, neural control, expert control, etc. and the combination of them or with traditional control theories have also been widely used in temperature intelligent control, improving the The automatic control capability of the system enhances the temperature and robustness of the system.

近十年来，深度学习在理论、算法和应用等方面发展迅速、进步显著。区别于传统的浅层学习，深度学习通过构建具有多个隐层的机器学习模型和大量的训练数据，学习更有用的特征，发掘数据的隐藏模式，最终提升预测的准确性。目前，深度学习已被成功应用到诸多显示领域中，如图像和视觉处理、语音和语言处理、信息安全和棋类比赛。然而在循环冷却水系统的相关领域中，却没有深度学习的应用出现。In the past ten years, deep learning has developed rapidly and made remarkable progress in terms of theory, algorithm and application. Different from traditional shallow learning, deep learning builds a machine learning model with multiple hidden layers and a large amount of training data to learn more useful features, discover hidden patterns of data, and ultimately improve the accuracy of predictions. At present, deep learning has been successfully applied to many display fields, such as image and visual processing, speech and language processing, information security and chess games. However, in the related field of circulating cooling water system, there is no application of deep learning.

发明内容Contents of the invention

针对上述问题，本发明提供了一种循环冷却水系统工艺介质多温度目标设定值切换多参数预测控制算法，使循环冷却水工艺介质温度节能控制系统能够实时获得工业生产现场各换热器内工艺介质维持在各自温度设定范围内需要设定的循环冷却给回水压差，为待冷却生产区精确降温，在高精度的同时还能最大限度节能。In view of the above problems, the present invention provides a multi-temperature target setting value switching multi-parameter predictive control algorithm for the process medium of the circulating cooling water system, so that the energy-saving control system of the process medium temperature of the circulating cooling water can obtain real-time To maintain the process medium within the respective temperature setting range, it is necessary to set the pressure difference between the circulating cooling water supply and the return water, so as to accurately cool down the production area to be cooled, and save energy to the greatest extent while maintaining high precision.

为达到上述目的，本发明采用的具体技术方案如下：In order to achieve the above object, the concrete technical scheme that the present invention adopts is as follows:

一种循环冷却水系统工艺介质多温度目标设定值切换多参数预测控制算法，其中，循环冷却水系统包括沿着循环水路设置的N’个冷却塔、吸水池、给水泵机组、出水管组、M’个生产区换热器以及回水管组；在冷却塔内设置有冷却池；A multi-parameter predictive control algorithm for multi-temperature target setting value switching of process medium in a circulating cooling water system, wherein the circulating cooling water system includes N' cooling towers, water suction pools, feed water pump units, and outlet pipe groups arranged along the circulating water path , M' production area heat exchangers and return water pipe groups; a cooling pool is provided in the cooling tower;

所述出水管组包括N’根吸水池进水管、L根吸水池出水管、给水总管、M’根给水支管；The water outlet pipe group includes N' suction pool inlet pipes, L water suction pool outlet pipes, water supply main pipe, M' water supply branch pipes;

所述回水管组包括M’根回水支管、回水总管、N’根上塔回水管；The return pipe group includes M' backwater branch pipe, backwater main pipe, N' upper tower return pipe;

任一所述冷却塔经一根对应的吸水池进水管与所述吸水池连接，所述吸水池经L根并列的吸水池出水管与所述给水总管连接，所述给水总管经M’根给水支管向M’个生产区换热器一一对应供水；Any of the cooling towers is connected to the water-absorbing pool through a corresponding water-absorbing pool inlet pipe, and the water-absorbing pool is connected to the water supply main pipe through L parallel water-absorbing pool outlet pipes, and the water supply main pipe is connected to the water supply main pipe through M' The branch water supply pipe supplies water to the heat exchangers in M' production areas in one-to-one correspondence;

在所述吸水池出水管上并联设置有给水泵；A water feed pump is arranged in parallel on the water outlet pipe of the water absorption pool;

任一所述生产区换热器经一根对应的回水支管与所述回水总管连接，所述回水总管经并列的N’根上塔回水管与N’个冷却塔一一对应连接；Any one of the heat exchangers in the production area is connected to the return water main pipe through a corresponding return water branch pipe, and the return water main pipe is connected to N' cooling towers one by one through parallel N' upper tower return water pipes;

M’个所述生产区换热器内均设置有工艺介质温度传感器，该工艺介质温度传感器用于获取各种工艺介质实时温度检测值；Process medium temperature sensors are all arranged in the heat exchangers of M' described production areas, and the process medium temperature sensors are used to obtain real-time temperature detection values of various process mediums;

在所述给水总管上设置有冷却给水温度传感器和冷却给水压力传感器，所述冷却给水温度传感器用于获取冷却给水温度检测值，所述冷却给水压力传感器用于获取冷却给水压力检测值；A cooling feed water temperature sensor and a cooling feed water pressure sensor are arranged on the feed water main pipe, the cooling feed water temperature sensor is used to obtain the detection value of the cooling feed water temperature, and the cooling feed water pressure sensor is used to obtain the detection value of the cooling feed water pressure;

在所述回水总管上设置有冷却回水温度传感器和冷却回水压力传感器，所述冷却回水温度传感器用于获取冷却回水温度检测值，所述冷却回水压力传感器用于获取冷却回水压力检测值；A cooling return water temperature sensor and a cooling return water pressure sensor are arranged on the return water main pipe, the cooling return water temperature sensor is used to obtain the detection value of the cooling return water temperature, and the cooling return water pressure sensor is used to obtain the cooling return water Water pressure detection value;

在N’根所述上塔回水管上分别设置有一个上塔阀；An upper tower valve is respectively arranged on the upper tower return pipes of the N' roots;

其关键技术在于：循环冷却水系统工艺介质多温度目标设定值切换多参数预测控制算法的步骤为：The key technology lies in: the steps of multi-parameter predictive control algorithm switching multi-parameter predictive control algorithm for the multi-temperature target set value of the process medium in the circulating cooling water system are as follows:

S1：对M’个换热器进行编号，并获取循环冷却水系统在a个采样周期内运行产生的历史数据，并将获取的历史数据作为换热器内工艺介质温度预测模型的训练数据；S1: Number the M' heat exchangers, and obtain the historical data generated by the operation of the circulating cooling water system in a sampling period, and use the acquired historical data as the training data for the temperature prediction model of the process medium in the heat exchanger;

S2：从历史数据中根据筛选条件筛选出特征变量，将换热器的历史数据中的工艺介质温度检测值、工艺介质温度偏差、工艺介质温度偏差变化率、冷却给水温度检测值、冷却给回水压差检测值作为输入数据，并进行数据归一化处理后得到归一数据集，并将该归一数据集划分为训练样本集和测试样本集；S2: Select the characteristic variables from the historical data according to the screening conditions, and collect the process medium temperature detection value, process medium temperature deviation, process medium temperature deviation change rate, cooling feed water temperature detection value, and cooling feed return in the historical data of the heat exchanger. The detected value of the water pressure difference is used as the input data, and the normalized data set is obtained after data normalization processing, and the normalized data set is divided into a training sample set and a test sample set;

S3：基于堆叠自动编码器，对训练样本集进行逐层贪婪无监督预训练，得到深层神经网络初始化的输入层与隐含层的权值矩阵W、输入层与隐含层阈值矩阵B；S3: Based on the stacked autoencoder, perform layer-by-layer greedy unsupervised pre-training on the training sample set, and obtain the weight matrix W of the input layer and hidden layer initialized by the deep neural network, and the threshold matrix B of the input layer and hidden layer;

S4：进行参数微调：对深层神经网络初始化的输入层与隐含层的权值矩阵W、输入层与隐含层阈值矩阵B进行微调，直到迭代次数达到迭代次数最大值为止，得到基于堆叠自动编码器的初始工艺介质温度预测模型；S4: Perform parameter fine-tuning: fine-tune the weight matrix W of the input layer and hidden layer initialized by the deep neural network, and the threshold matrix B of the input layer and hidden layer until the number of iterations reaches the maximum number of iterations, and the stack-based automatic The initial process medium temperature prediction model of the encoder;

S5：使用测试样本数据集对步骤S4得到的初始工艺介质温度预测模型进行测试，得到基于堆叠自动编码器的工艺介质温度预测模型；S5: Use the test sample data set to test the initial process medium temperature prediction model obtained in step S4, and obtain a process medium temperature prediction model based on stacked autoencoders;

S6：对步骤S5得到的工艺介质温度预测模型进行评估；S6: Evaluate the process medium temperature prediction model obtained in step S5;

S7：获取循环冷却水系统当前状态下的当前数据，确定当前状态下的工艺介质最不利点，确定换热器控制对象，并结合工艺介质温度预测模型，得到对应的冷却给回水压差设定值和工艺介质温度预测值。S7: Obtain the current data of the circulating cooling water system in the current state, determine the most unfavorable point of the process medium in the current state, determine the control object of the heat exchanger, and combine the temperature prediction model of the process medium to obtain the corresponding cooling water supply and return water pressure difference setting Fixed value and predicted value of process medium temperature.

通过上述设计，通过获取历史数据，可以有效表征工艺介质温度值、工业循环冷却给水温度与冷却给水压力设定值之间的复杂函数，快速准确的对工业生产现场工艺介质温度进行预测控制，同时该算法有良好的泛化能力，对具有不同温度变化特征的工艺介质预测适应能力强。与人工经验计算调节相比，极大提高了工艺介质温度控制的准确性，节省了计算时间开销，有助于工业生产现场管理人员实时掌握各换热器内工艺介质温度变化趋势。Through the above design, by obtaining historical data, it is possible to effectively characterize the complex function between the temperature value of the process medium, the temperature of the industrial circulating cooling feed water, and the set value of the pressure of the cooling feed water, and quickly and accurately predict and control the temperature of the process medium on the industrial production site. The algorithm has good generalization ability and strong adaptability to process medium prediction with different temperature change characteristics. Compared with manual empirical calculation and adjustment, it greatly improves the accuracy of process medium temperature control, saves calculation time and expenses, and helps industrial production site managers to grasp the temperature change trend of process medium in each heat exchanger in real time.

循环冷却水系统检测得到的冷却给水压力检测值与冷却回水压力检测值作差后得到冷却给回水压差检测值；该冷却给回水压差检测值与冷却给回水压差设定值作差后得到的冷却给回水压差偏差值；所述给回水压差内环PID控制器根据冷却给回水压差偏差值来调节所有所述上塔阀的开度，从而改变给水泵机组的出口水流量。The difference between the detected value of the cooling feed water pressure detected by the circulating cooling water system and the detected value of the cooling return water pressure is used to obtain the detected value of the differential pressure of the cooling feed water; The deviation value of the cooling feed-back water pressure difference obtained after making a difference; the inner ring PID controller of the feed-back water pressure difference adjusts the openings of all the upper tower valves according to the difference value of the cooling feed-back water pressure difference, thereby changing The outlet water flow rate of the feedwater pump unit.

进一步的，所述筛选条件是工艺介质温度处于安全且节能的温度值区间内的历史数据，所述安全且节能的温度值区间在工艺介质温度阈值区间内。Further, the screening condition is historical data that the temperature of the process medium is within a safe and energy-saving temperature value interval, and the safe and energy-saving temperature value interval is within the threshold value interval of the process medium temperature.

安全且节能的温度值区间和工艺介质温度阈值区间之间的差值，根据技术人员的历史经验技术进行归纳后设定。The difference between the safe and energy-saving temperature range and the process medium temperature threshold range is set after induction based on the technical personnel's historical experience.

再进一步的，步骤S3中，基于堆栈自动编码器的，对训练样本集进行逐层贪婪无监督预训练时，将训练样本集分成P组小批量训练样本，依次进行训练，并采用Dropout技术，随机选取部分神经元暂停工作，依次迭代，逐层训练，得到深层神经网络初始化的输入层与隐含层的权值矩阵W、输入层与隐含层阈值矩阵B。Further, in step S3, when performing layer-by-layer greedy unsupervised pre-training on the training sample set based on the stack autoencoder, the training sample set is divided into P groups of small batch training samples, and the training is performed sequentially, and the Dropout technology is used. Randomly select some neurons to suspend work, iterate sequentially, and train layer by layer to obtain the weight matrix W of the input layer and hidden layer and the threshold matrix B of the input layer and hidden layer initialized by the deep neural network.

具体步骤：Specific steps:

步骤一：将工业生产现场循环冷却水系统运行产生的历史数据作为换热器内工艺介质温度预测模型的训练数据；Step 1: Use the historical data generated by the operation of the circulating cooling water system on the industrial production site as the training data for the temperature prediction model of the process medium in the heat exchanger;

步骤二：筛选出步骤一中历史数据的特征变量，将编号I换热器内工艺介质温度偏差、编号I换热器内工艺介质温度偏差变化率、冷却给水温度检测值、冷却给回水压差检测值作为输入数据，并进行数据归一化处理后，划分为训练样本集和测试样本集两部分；Step 2: Filter out the characteristic variables of the historical data in step 1, and compare the temperature deviation of the process medium in the number I heat exchanger, the change rate of the temperature deviation of the process medium in the number I heat exchanger, the detection value of the cooling feed water temperature, and the cooling water return water pressure The difference detection value is used as input data, and after data normalization processing, it is divided into two parts: training sample set and test sample set;

步骤三：设定隐含层层数以及每层隐含层神经元的个数；Step 3: Set the number of hidden layers and the number of neurons in each hidden layer;

步骤四：逐层贪婪无监督预训练，依次迭代，逐层训练，得到深层神经网络初始化的输入层与隐含层的权值矩阵W、输入层与隐含层阈值矩阵B；Step 4: Layer-by-layer greedy unsupervised pre-training, successive iterations, layer-by-layer training, to obtain the weight matrix W of the input layer and hidden layer initialized by the deep neural network, and the threshold matrix B of the input layer and hidden layer;

步骤五：参数微调，对初深层神经网络初始化的输入层与隐含层的权值矩阵W、输入层与隐含层阈值矩阵B进行微调，直到迭代次数达到设定最大值为止。工艺介质温度预测控制算法包括基于堆栈自动编码器的换热器内工艺介质温度预测模型离线训练和在线应用给出循环冷却水系统冷却给回水压差设定值：堆栈自动编码器属于深层神经网络，与传统的浅层神经网络相比，堆栈自动编码器有效地解决了传统神经网络参数随机初始化导致的一系列问题，可以有效挖掘各数据隐含关系，极大地提高了工业生产现场工艺介质温度预测控制准确性。其中I为0-50的整数，当确定最不利点后，进而确定I值大小，来调取的工艺介质温度预测模型，再结合最不利点换热器的现场运行数据，得到冷却给回水压差设定值。Step 5: Parameter fine-tuning. Fine-tune the weight matrix W of the input layer and hidden layer and the threshold matrix B of the input layer and hidden layer initialized in the initial deep neural network until the number of iterations reaches the set maximum value. The process medium temperature predictive control algorithm includes the offline training and online application of the process medium temperature prediction model in the heat exchanger based on the stack autoencoder to give the set value of the cooling water supply and return water pressure difference in the circulating cooling water system: the stack autoencoder belongs to the deep neural network Compared with the traditional shallow neural network, the stack autoencoder effectively solves a series of problems caused by the random initialization of the traditional neural network parameters, can effectively mine the hidden relationship of each data, and greatly improves the industrial production site process medium. Temperature predictive control accuracy. Among them, I is an integer of 0-50. When the most unfavorable point is determined, the value of I is further determined to obtain the temperature prediction model of the process medium, and combined with the on-site operation data of the heat exchanger at the most unfavorable point, the cooling water supply and return water is obtained. differential pressure setting.

工艺介质温度预测控制算法包括基于堆栈自动编码器的换热器内工艺介质温度预测模型离线训练和在线应用给出循环冷却水系统冷却给回水压差设定值：堆栈自动编码器属于深层神经网络，与传统的浅层神经网络相比，堆栈自动编码器有效地解决了传统神经网络参数随机初始化导致的一系列问题，可以有效挖掘各数据隐含关系，极大地提高了工业生产现场工艺介质温度预测控制准确性。The process medium temperature predictive control algorithm includes the offline training and online application of the process medium temperature prediction model in the heat exchanger based on the stack autoencoder to give the set value of the cooling water supply and return water pressure difference in the circulating cooling water system: the stack autoencoder belongs to the deep neural network Compared with the traditional shallow neural network, the stack autoencoder effectively solves a series of problems caused by the random initialization of the traditional neural network parameters, can effectively mine the hidden relationship of each data, and greatly improves the industrial production site process medium. Temperature predictive control accuracy.

再进一步的，步骤S6工艺介质温度预测模型进行评估时；Still further, when evaluating the process medium temperature prediction model in step S6;

使用测试样本数据集测试训练之后的改进堆栈自动编码器，采用平均百分比误差(MAPE)作为衡量改进堆栈自动编码器评估性能的标准，表达式为：Use the test sample data set to test the improved stack autoencoder after training, and use the average percentage error (MAPE) as the standard to measure the evaluation performance of the improved stack autoencoder. The expression is:

式中：yⁱ、

分别为第i个样本循环冷却给回水压差的实际值和通过SAE评估得到的预测值。In the formula: y ⁱ ,

Respectively, the actual value of the circulating cooling feed-backwater pressure difference of the i-th sample and the predicted value obtained through SAE evaluation.

再进一步描述，步骤S4中，参数微调时，采用自上而下的小批量RMSProp优化方法对深层神经网络初始化的输入层与隐含层的权值矩阵W、输入层与隐含层阈值矩阵B进行微调，具体步骤为：To further describe, in step S4, when parameters are fine-tuned, the top-down small batch RMSProp optimization method is used to initialize the weight matrix W of the input layer and hidden layer of the deep neural network, and the threshold matrix B of the input layer and hidden layer For fine-tuning, the specific steps are:

S41，设置全局学习率l、衰减速率ρ，初始化累计变量r1＝0，r2＝0。S41, setting the global learning rate l and the decay rate ρ, and initializing the cumulative variables r1=0, r2=0.

S42，从训练集中选取包含m个样本的小批量数据集，根据误差损失函数，计算梯度：S42, select a small batch data set containing m samples from the training set, and calculate the gradient according to the error loss function:

S43，计算累计平方梯度，如式(9)所示：S43, calculate the cumulative square gradient, as shown in formula (9):

式中：⊙为逐元素乘积符号。In the formula: ⊙ is the element-wise product symbol.

S44，分别更新权重和阈值参数：S44, updating weight and threshold parameters respectively:

S45，当迭代次数达到要求时，停止运算，否则返回第S42步继续执行计算。S45, when the number of iterations reaches the requirement, stop the operation, otherwise return to step S42 to continue the calculation.

再进一步描述，步骤S7的具体步骤为：To further describe, the specific steps of step S7 are:

S71：获取循环冷却水系统当前状态下的当前数据；S71: Obtain current data in the current state of the circulating cooling water system;

S72：根据当前数据，确定当前状态下的工艺介质最不利点，S72: According to the current data, determine the most unfavorable point of the process medium under the current state,

S73：确定该工艺介质最不利点对应的换热器j；S73: Determine the heat exchanger j corresponding to the most unfavorable point of the process medium;

S74：获取的换热器j的特征变量；S74: The acquired characteristic variables of the heat exchanger j;

S75：对换热器j的特征变量进行数据归一化处理；S75: Perform data normalization processing on the characteristic variables of the heat exchanger j;

S76：结合工艺介质温度预测模型和步骤S75得到的数据，确定冷却给回水压差设定值和工艺介质温度预测值。S76: Combining the temperature prediction model of the process medium with the data obtained in step S75, determine the set value of the cooling feed-backwater pressure difference and the predicted value of the temperature of the process medium.

再进一步描述，所述特征变量包括任意一个换热器内的工艺介质温度偏差值、工艺介质温度偏差变化率、工艺介质实时温度检测值以及循环冷却水系统中的冷却给水温度检测值、冷却给回水压差检测值。To further describe, the characteristic variables include the temperature deviation value of the process medium in any heat exchanger, the rate of change of the temperature deviation of the process medium, the real-time temperature detection value of the process medium, the temperature detection value of the cooling feed water in the circulating cooling water system, the cooling feed Return water pressure differential detection value.

工艺介质温度偏差值等于对应所述工艺介质实时温度检测值与对应所述工艺介质温度设定值的差值；The temperature deviation value of the process medium is equal to the difference between the real-time temperature detection value corresponding to the process medium and the set value corresponding to the process medium temperature;

所述工艺介质温度偏差变化率为对应所述工艺介质相邻两个检测时段温度变化值与上一检测时段的比值。The temperature deviation change rate of the process medium corresponds to the ratio of the temperature change value of the process medium in two adjacent detection periods to the previous detection period.

再进一步描述，步骤S72确定当前状态下的工艺介质最不利点的步骤为：To further describe, step S72 determines the steps of the most unfavorable point of the process medium in the current state as follows:

S721：初始化，设M’个工艺介质温度偏差值组成一个差值小组，共计M’个工艺介质温度偏差值，令W_k＝M’；k＝1S721: Initialize, set M' process medium temperature deviation values to form a difference group, a total of M' process medium temperature deviation values, set W _k = M'; k = 1

S722：令W_k+1＝W_k+X，使W_k+1可以被M’整除，X为填充的差值无线大的空位；且X等于0～M’-1；S722: Let W _k+1 =W _k +X, so that W _k+1 can be divisible by M', and X is a gap filled with an infinitely large difference; and X is equal to 0~M'-1;

S723：计算W_k+2＝W_k+1/M’S723: Calculate W _k+2 =W _k+1 /M'

S724：从W_k+2组中，采用交叉比较法，从每一组的M’个工艺介质温度偏差值中找出最小值，得到W_k+2个工艺介质温度偏差值；S724: From the W _k+2 group, use the cross-comparison method to find the minimum value from the M' process medium temperature deviation values in each group, and obtain W _k+2 process medium temperature deviation values;

S725：判断W_k+2是否等于1；若是，将该工艺介质温度偏差值作为工艺介质温度最不利点；否则，令k＝k+2；返回步骤S722。S725: Determine whether W _k+2 is equal to 1; if so, use the process medium temperature deviation value as the most unfavorable point of process medium temperature; otherwise, set k=k+2; return to step S722.

采用上述方案，寻找大数量的工艺介质温度偏差最小值，以确定工艺介质温度最不利点，注重工艺介质最不利点选择的实时性和准确性。Using the above scheme, find the minimum value of the temperature deviation of a large number of process media to determine the most unfavorable point of the process medium temperature, and pay attention to the real-time and accuracy of the selection of the most unfavorable point of the process medium.

本发明的有益效果：使循环冷却水工艺介质温度节能控制系统能够实时获得工业生产现场各换热器内工艺介质维持在各自温度设定范围内需要设定的循环冷却给回水压差，为待冷却生产区精确降温，在高精度的同时还能最大限度节能。The beneficial effect of the present invention is to enable the energy-saving control system for the temperature of the circulating cooling water process medium to obtain in real time the pressure difference of the circulating cooling feed and return water that needs to be set to maintain the process medium in each heat exchanger in the industrial production site within the respective temperature setting range, for Precise cooling of the production area to be cooled can maximize energy saving while maintaining high precision.

附图说明Description of drawings

图1是化工厂A循环冷却水系统工艺图；Figure 1 is a process diagram of the circulating cooling water system of chemical plant A;

图2是化工厂A给水泵机组变频变压供水控制框图；Fig. 2 is a control block diagram of variable frequency and variable pressure water supply of the water supply pump unit of chemical plant A;

图3是化工厂A循环冷却水系统中水合成区换热器组示意图；Fig. 3 is a schematic diagram of the heat exchanger group in the hydration zone in the circulating cooling water system of chemical plant A;

图4是本发明工艺介质多温度目标设定值切换多参量预测控制算法流程图；Fig. 4 is a flow chart of a multi-parameter predictive control algorithm for switching between multi-temperature target set values of the process medium of the present invention;

图5是本发明自动编码器结构图；Fig. 5 is a structural diagram of the automatic encoder of the present invention;

图6是本发明堆叠自编码器架构；Fig. 6 is the stacked self-encoder architecture of the present invention;

图7是本发明Dropout自编码器架构；Fig. 7 is the Dropout self-encoder framework of the present invention;

图8是本发明工艺介质温度最不利点寻找流程图。Fig. 8 is a flow chart of finding the most unfavorable point of process medium temperature in the present invention.

具体实施方式Detailed ways

下面结合附图对本发明的具体实施方式以及工作原理作进一步详细说明。The specific implementation manner and working principle of the present invention will be further described in detail below in conjunction with the accompanying drawings.

在本实施例中，以化工厂A为例，循环冷却水节能控制。。从图1-3可以看出，一种循环冷却水系统工艺介质多温度目标设定值切换多参数预测控制算法，其中，循环冷却水系统包括沿着循环水路设置的N’个冷却塔、吸水池、给水泵机组、出水管组、M’个生产区换热器以及回水管组；在冷却塔内设置有冷却池；在本实施例中，M’为正整数，有三个换热器组：合成区换热器组、尿素区换热器组、动力区换热器组。In this embodiment, taking chemical plant A as an example, the energy-saving control of circulating cooling water is carried out. . It can be seen from Figure 1-3 that a multi-parameter predictive control algorithm for multi-temperature target set value switching of process medium in a circulating cooling water system, wherein the circulating cooling water system includes N' cooling towers arranged along the circulating waterway, water suction Pool, feed water pump unit, outlet pipe group, M' heat exchangers in the production area and return water pipe group; a cooling pool is provided in the cooling tower; in this embodiment, M' is a positive integer, and there are three heat exchanger groups : Heat exchanger group in synthesis area, heat exchanger group in urea area, heat exchanger group in power area.

所述出水管组包括N’根吸水池进水管、L根吸水池出水管、给水总管、M’根给水支管；M’＝50。在本实施例中，换热器数量M’＝50。其中，50个换热器分成三个换热器组，分别为合成区换热器组、尿素去换热器组、动力去换热器组。三个换热器组共计50个换热器。The water outlet pipe group includes N' suction pool inlet pipes, L water suction pool outlet pipes, water supply main pipe, M' water supply branch pipes; M'=50. In this embodiment, the number of heat exchangers M'=50. Among them, 50 heat exchangers are divided into three heat exchanger groups, namely the synthesis zone heat exchanger group, the urea heat exchanger group, and the power heat exchanger group. The three heat exchanger groups have a total of 50 heat exchangers.

在本实施例中，所述回水管组包括M’根回水支管、回水总管、N’根上塔回水管；In this embodiment, the return pipe group includes M' return water branch pipes, return water main pipes, and N' upper tower return pipes;

在本实施例中，任一所述冷却塔经一根对应的吸水池进水管与所述吸水池连接，所述吸水池经L根并列的吸水池出水管与所述给水总管连接，所述给水总管经M’根给水支管向M’个生产区换热器一一对应供水；In this embodiment, any one of the cooling towers is connected to the water-absorbing pool through a corresponding water-absorbing pool inlet pipe, and the water-absorbing pool is connected to the water supply main pipe through L parallel water-absorbing pool outlet pipes. The main water supply pipe supplies water to the heat exchangers in M' production areas one by one through M' water supply branch pipes;

在所述吸水池出水管上并联设置有给水泵；A water feed pump is arranged in parallel on the water outlet pipe of the water absorption pool;

在本实施例中，任一所述生产区换热器经一根对应的回水支管与所述回水总管连接，所述回水总管经并列的N’根上塔回水管与N’个冷却塔一一对应连接；In this embodiment, any one of the heat exchangers in the production area is connected to the return water main pipe through a corresponding return water branch pipe, and the return water main pipe is connected to N' cooling water pipes through parallel N' upper tower return water pipes. One-to-one connection of towers;

在本实施例中，M’个所述生产区换热器内均设置有工艺介质温度传感器，该工艺介质温度传感器用于获取各种工艺介质实时温度检测值；在所述给水总管上设置有冷却给水温度传感器和冷却给水压力传感器，所述冷却给水温度传感器用于获取冷却给水温度检测值，所述冷却给水压力传感器用于获取冷却给水压力检测值；在所述回水总管上设置有冷却回水温度传感器和冷却回水压力传感器，所述冷却回水温度传感器用于获取冷却回水温度检测值，所述冷却回水压力传感器用于获取冷却回水压力检测值；In this embodiment, the M' heat exchangers in the production area are all equipped with process medium temperature sensors, which are used to obtain real-time temperature detection values of various process mediums; A cooling feed water temperature sensor and a cooling feed water pressure sensor, the cooling feed water temperature sensor is used to obtain the detection value of the cooling feed water temperature, and the cooling feed water pressure sensor is used to obtain the detection value of the cooling feed water pressure; A return water temperature sensor and a cooling return water pressure sensor, the cooling return water temperature sensor is used to obtain the detection value of the cooling return water temperature, and the cooling return water pressure sensor is used to obtain the detection value of the cooling return water pressure;

在本实施例中，在N’根所述上塔回水管上分别设置有一个上塔阀。In this embodiment, an upper tower valve is respectively arranged on the N' return water pipes of the upper tower.

其关键技术在于：结合图4可以看出，循环冷却水系统工艺介质多温度目标设定值切换多参数预测控制算法的步骤为：Its key technology is: combined with Figure 4, it can be seen that the steps of the multi-parameter predictive control algorithm for switching the multi-temperature target set value of the process medium in the circulating cooling water system are as follows:

S1：对M’个换热器进行编号，并获取循环冷却水系统在a个采样周期内运行产生的历史数据，并将获取的历史数据作为换热器内工艺介质温度预测模型的训练数据；S1: Number the M' heat exchangers, and obtain the historical data generated by the operation of the circulating cooling water system in a sampling period, and use the acquired historical data as the training data for the temperature prediction model of the process medium in the heat exchanger;

S2：从历史数据中根据筛选条件筛选出特征变量，将换热器的历史数据中的工艺介质温度检测值、工艺介质温度偏差、工艺介质温度偏差变化率、冷却给水温度检测值、冷却给回水压差检测值作为输入数据，并进行数据归一化处理后得到归一数据集，并将该归一数据集划分为训练样本集和测试样本集；S2: Select the characteristic variables from the historical data according to the screening conditions, and collect the process medium temperature detection value, process medium temperature deviation, process medium temperature deviation change rate, cooling feed water temperature detection value, and cooling feed return in the historical data of the heat exchanger. The detected value of the water pressure difference is used as the input data, and the normalized data set is obtained after data normalization processing, and the normalized data set is divided into a training sample set and a test sample set;

S3：基于堆叠自动编码器，对训练样本集进行逐层贪婪无监督预训练，得到深层神经网络初始化的输入层与隐含层的权值矩阵W、输入层与隐含层阈值矩阵B；S3: Based on the stacked autoencoder, perform layer-by-layer greedy unsupervised pre-training on the training sample set, and obtain the weight matrix W of the input layer and hidden layer initialized by the deep neural network, and the threshold matrix B of the input layer and hidden layer;

S4：进行参数微调：对深层神经网络初始化的输入层与隐含层的权值矩阵W、输入层与隐含层阈值矩阵B进行微调，直到迭代次数达到迭代次数最大值为止，得到基于堆叠自动编码器的初始工艺介质温度预测模型；S4: Perform parameter fine-tuning: fine-tune the weight matrix W of the input layer and hidden layer initialized by the deep neural network, and the threshold matrix B of the input layer and hidden layer until the number of iterations reaches the maximum number of iterations, and the stack-based automatic The initial process medium temperature prediction model of the encoder;

S5：使用测试样本数据集对步骤S4得到的初始工艺介质温度预测模型进行测试，得到基于堆叠自动编码器的工艺介质温度预测模型；S5: Use the test sample data set to test the initial process medium temperature prediction model obtained in step S4, and obtain a process medium temperature prediction model based on stacked autoencoders;

S6：对步骤S5得到的工艺介质温度预测模型进行评估；S6: Evaluate the process medium temperature prediction model obtained in step S5;

S7：获取循环冷却水系统当前状态下的当前数据，确定当前状态下的工艺介质最不利点，确定换热器控制对象，并结合工艺介质温度预测模型，得到对应的冷却给回水压差设定值P_Δs和工艺介质温度预测值T_yi。S7: Obtain the current data of the circulating cooling water system in the current state, determine the most unfavorable point of the process medium in the current state, determine the control object of the heat exchanger, and combine the temperature prediction model of the process medium to obtain the corresponding cooling water supply and return water pressure difference setting Fixed value P _Δs and predicted value T _yi of process medium temperature.

进一步的，所述筛选条件是工艺介质温度处于安全且节能的温度值区间内的历史数据，所述安全且节能的温度值区间在工艺介质温度阈值区间内。Further, the screening condition is historical data that the temperature of the process medium is within a safe and energy-saving temperature value interval, and the safe and energy-saving temperature value interval is within the threshold value interval of the process medium temperature.

安全且节能的温度值区间和工艺介质温度阈值区间之间的差值，根据技术人员的历史经验技术进行归纳后设定。The difference between the safe and energy-saving temperature range and the process medium temperature threshold range is set after induction based on the technical personnel's historical experience.

在本实施例中，步骤S3中，基于堆栈自动编码器的，对训练样本集进行逐层贪婪无监督预训练时，将训练样本集分成P组小批量训练样本，依次进行训练，并采用Dropout技术，随机选取部分神经元暂停工作，依次迭代，逐层训练，得到深层神经网络初始化的输入层与隐含层的权值矩阵W、输入层与隐含层阈值矩阵B。In this embodiment, in step S3, when performing layer-by-layer greedy unsupervised pre-training on the training sample set based on the stack autoencoder, the training sample set is divided into P groups of small-batch training samples, and the training is performed sequentially, and Dropout is used. The technology randomly selects some neurons to suspend work, iterates sequentially, and trains layer by layer to obtain the weight matrix W of the input layer and hidden layer and the threshold matrix B of the input layer and hidden layer initialized by the deep neural network.

具体步骤：Specific steps:

步骤一：将工业生产现场循环冷却水系统运行产生的历史数据作为换热器内工艺介质温度预测模型的训练数据；Step 1: Use the historical data generated by the operation of the circulating cooling water system on the industrial production site as the training data for the temperature prediction model of the process medium in the heat exchanger;

步骤二：筛选出步骤一中历史数据的特征变量，将编号I换热器内工艺介质温度偏差、编号I换热器内工艺介质温度偏差变化率、冷却给水温度检测值、冷却给回水压差检测值作为输入数据，并进行数据归一化处理后，划分为训练样本集和测试样本集两部分；Step 2: Filter out the characteristic variables of the historical data in step 1, and compare the temperature deviation of the process medium in the number I heat exchanger, the change rate of the temperature deviation of the process medium in the number I heat exchanger, the detection value of the cooling feed water temperature, and the cooling water return water pressure The difference detection value is used as input data, and after data normalization processing, it is divided into two parts: training sample set and test sample set;

步骤三：设定隐含层层数以及每层隐含层神经元的个数；Step 3: Set the number of hidden layers and the number of neurons in each hidden layer;

步骤四：逐层贪婪无监督预训练，依次迭代，逐层训练，得到深层神经网络初始化的输入层与隐含层的权值矩阵W、输入层与隐含层阈值矩阵B；Step 4: Layer-by-layer greedy unsupervised pre-training, successive iterations, layer-by-layer training, to obtain the weight matrix W of the input layer and hidden layer initialized by the deep neural network, and the threshold matrix B of the input layer and hidden layer;

步骤五：参数微调，对初深层神经网络初始化的输入层与隐含层的权值矩阵W、输入层与隐含层阈值矩阵B进行微调，直到迭代次数达到设定最大值为止。工艺介质温度预测控制算法包括基于堆栈自动编码器的换热器内工艺介质温度预测模型离线训练和在线应用给出循环冷却水系统冷却给回水压差设定值：堆栈自动编码器属于深层神经网络，与传统的浅层神经网络相比，堆栈自动编码器有效地解决了传统神经网络参数随机初始化导致的一系列问题，可以有效挖掘各数据隐含关系，极大地提高了工业生产现场工艺介质温度预测控制准确性。Step 5: Parameter fine-tuning. Fine-tune the weight matrix W of the input layer and hidden layer and the threshold matrix B of the input layer and hidden layer initialized in the initial deep neural network until the number of iterations reaches the set maximum value. The process medium temperature predictive control algorithm includes the offline training and online application of the process medium temperature prediction model in the heat exchanger based on the stack autoencoder to give the set value of the cooling water supply and return water pressure difference in the circulating cooling water system: the stack autoencoder belongs to the deep neural network Compared with the traditional shallow neural network, the stack autoencoder effectively solves a series of problems caused by the random initialization of the traditional neural network parameters, can effectively mine the hidden relationship of each data, and greatly improves the industrial production site process medium. Temperature predictive control accuracy.

SAE是一种典型的深层神经网络，其基本构成单元是自动编码器(antoencoder,AE)，其网络结构如图5所示。自动编码器(autoencoder，AE)网络结构如图5所示，由编码器和解码器组成：通过编码器将输入向量映射为隐含层中的特征向量，然后通过解码器将特征相量重构为原来的输入向量。SAE is a typical deep neural network, and its basic constituent unit is an automatic encoder (antoencoder, AE), and its network structure is shown in Figure 5. The autoencoder (AE) network structure is shown in Figure 5, which consists of an encoder and a decoder: the input vector is mapped to the feature vector in the hidden layer through the encoder, and then the feature phasor is reconstructed through the decoder is the original input vector.

在本实施例中，具体步骤为：给定一个输入样本集合X＝{xⁱ|1≤i≤N}，其中，N为样本总个数，xⁱ为样本集中的第i个训练样本，维数为n。设H＝{hⁱ|1≤i≤N}为隐含层特征向量集合，hⁱ为第i个样本对应的特征向量，维数为m，则X与H的编码关系为：In this embodiment, the specific steps are: given an input sample set X={x ⁱ |1≤i≤N}, where N is the total number of samples, x ⁱ is the i-th training sample in the sample set, The number of dimensions is n. Suppose H={h ⁱ |1≤i≤N} is the hidden layer feature vector set, h ⁱ is the feature vector corresponding to the i-th sample, and the dimension is m, then the encoding relationship between X and H is:

H＝s_f(WX+B)H=s _f (WX+B)

式中：W为输入层与隐含层的权值矩阵；B为输入层与隐含层阈值矩阵；sf为编码器的神经元激活函数，通常采用sigmoid函数，其具有良好的特征辨识度：In the formula: W is the weight matrix of the input layer and the hidden layer; B is the threshold matrix of the input layer and the hidden layer; sf is the neuron activation function of the encoder, usually using the sigmoid function, which has good feature recognition:

s_f(z)＝1/(1+exp(-z))s _f (z)=1/(1+exp(-z))

式中：z为输入向量。In the formula: z is the input vector.

解码器是编码器的逆运算，以隐含层的特征向量作为输入向量，设

为输出向量集合，

为第i个样本对应的输出向量，维数为n，则解码器的表达式为：The decoder is the inverse operation of the encoder, with the feature vector of the hidden layer as the input vector, set

is the set of output vectors,

is the output vector corresponding to the i-th sample, and the dimension is n, then the expression of the decoder is:

式中：W′为隐含层与输出层的权值矩阵；B′为隐含层与输出层的阈值矩阵；sg为解码器的神经元激活函数。In the formula: W' is the weight matrix of the hidden layer and the output layer; B' is the threshold matrix of the hidden layer and the output layer; sg is the neuron activation function of the decoder.

自动编码器通过最小化输出向量与输入向量之间的重构误差来达到特征提取的目的，重构误差的公式如下：The autoencoder achieves the purpose of feature extraction by minimizing the reconstruction error between the output vector and the input vector. The formula for the reconstruction error is as follows:

利用梯度下降算法不断调整网络权值和阈值，降低重构误差，公式如下：Use the gradient descent algorithm to continuously adjust the network weights and thresholds to reduce the reconstruction error. The formula is as follows:

式中：l为学习率；

表示

对权值W求偏导；

表示

对阈值B求偏导。In the formula: l is the learning rate;

express

Find the partial derivative of the weight W;

express

Take the partial derivative with respect to the threshold B.

SAE是一种由AE栈式堆叠构成的深层神经网络模型，下层AE的输出将作为上层AE的输入。通过AE的堆叠实现特征的逐步抽象，最终形成更加紧凑、有用的特征，如图6所示为堆叠自编码器架构。W_n为第n-1层隐含层与第n层隐含层的权重矩阵，B_n为第n-1层隐含层与第n层隐含层的阈值矩阵。训练过程分为贪婪逐层无监督预训练和有监督微调两个步骤。贪婪逐层无监督预训练通过逐层训练得到网络的初始化权重和阈值，底层AE的输入为原始数据，隐含层输出数据作为上层AE的输入数据。SAE is a deep neural network model composed of AE stacks, and the output of the lower layer AE will be used as the input of the upper layer AE. The gradual abstraction of features is realized through the stacking of AE, and finally more compact and useful features are formed, as shown in Figure 6, which is the stacked autoencoder architecture. W _n is the weight matrix of the n-1 hidden layer and the n-th hidden layer, and B _n is the threshold matrix of the n-1 hidden layer and the n-th hidden layer. The training process is divided into two steps: greedy layer-by-layer unsupervised pre-training and supervised fine-tuning. Greedy layer-by-layer unsupervised pre-training obtains the initial weights and thresholds of the network through layer-by-layer training. The input of the bottom layer AE is the original data, and the output data of the hidden layer is used as the input data of the upper layer AE.

当分层预训练完成后，将隐含层堆叠，其输入数据与输出数据关系表示为：After the hierarchical pre-training is completed, the hidden layers are stacked, and the relationship between the input data and the output data is expressed as:

式中：f为激活函数，xⁱ为第i个训练样本的输入变量，W、B分别为逐层预训练得到的网络初始化权重和阈值，

为第i个样本的预测值。构造实际值与预测值的误差损失函数，公式如下：In the formula: f is the activation function, x ⁱ is the input variable of the i-th training sample, W and B are the network initialization weight and threshold obtained by layer-by-layer pre-training respectively,

is the predicted value of the i-th sample. Construct the error loss function of the actual value and the predicted value, the formula is as follows:

式中：N为样本总个数，yⁱ为第i个样本的实际值。通过自上而下的反向传播对整个网络权重和阈值进行微调，减小预测值与实际值的误差。In the formula: N is the total number of samples, and y ⁱ is the actual value of the i-th sample. Fine-tune the weights and thresholds of the entire network through top-down backpropagation to reduce the error between the predicted value and the actual value.

在本实施例中，步骤S6工艺介质温度预测模型进行评估时；In this embodiment, when evaluating the process medium temperature prediction model in step S6;

使用测试样本数据集测试训练之后的改进堆栈自动编码器，采用平均百分比误差(MAPE)作为衡量改进堆栈自动编码器评估性能的标准，表达式为：Use the test sample data set to test the improved stack autoencoder after training, and use the average percentage error (MAPE) as the standard to measure the evaluation performance of the improved stack autoencoder. The expression is:

式中：yⁱ、

分别为第i个样本循环冷却给回水压差的实际值和通过SAE评估得到的预测值。In the formula: y ⁱ ,

Respectively, the actual value of the circulating cooling feed-backwater pressure difference of the i-th sample and the predicted value obtained through SAE evaluation.

在本实施例中，步骤S4中，参数微调时，采用自上而下的小批量RMSProp优化方法对深层神经网络初始化的输入层与隐含层的权值矩阵W、输入层与隐含层阈值矩阵B进行微调，具体步骤为：In this embodiment, in step S4, when parameters are fine-tuned, the top-down small batch RMSProp optimization method is used to initialize the weight matrix W of the input layer and the hidden layer of the deep neural network, the threshold value of the input layer and the hidden layer Matrix B is fine-tuned, and the specific steps are:

S41，设置全局学习率l、衰减速率ρ，初始化累计变量r1＝0，r2＝0。S41, setting the global learning rate l and the decay rate ρ, and initializing the cumulative variables r1=0, r2=0.

S42，从训练集中选取包含m个样本的小批量数据集，根据误差损失函数，计算梯度：S42, select a small batch data set containing m samples from the training set, and calculate the gradient according to the error loss function:

S43，计算累计平方梯度，如式(9)所示：S43, calculate the cumulative square gradient, as shown in formula (9):

式中：⊙为逐元素乘积符号。In the formula: ⊙ is the element-wise product symbol.

S44，分别更新权重和阈值参数：S44, updating weight and threshold parameters respectively:

S45，当迭代次数达到要求时，停止运算，否则返回第S42步继续执行计算。S45, when the number of iterations reaches the requirement, stop the operation, otherwise return to step S42 to continue the calculation.

再进一步描述，步骤S7的具体步骤为：To further describe, the specific steps of step S7 are:

S71：获取循环冷却水系统当前状态下的当前数据；S71: Obtain current data in the current state of the circulating cooling water system;

S72：根据当前数据，确定当前状态下的工艺介质最不利点，S72: According to the current data, determine the most unfavorable point of the process medium under the current state,

S73：确定该工艺介质最不利点对应的换热器j；S73: Determine the heat exchanger j corresponding to the most unfavorable point of the process medium;

S74：获取的换热器j的特征变量；S74: The acquired characteristic variables of the heat exchanger j;

S75：对换热器j的特征变量进行数据归一化处理；S75: Perform data normalization processing on the characteristic variables of the heat exchanger j;

S76：结合工艺介质温度预测模型和步骤S75得到的数据，确定冷却给回水压差设定值和工艺介质温度预测值。S76: Combining the temperature prediction model of the process medium with the data obtained in step S75, determine the set value of the cooling feed-backwater pressure difference and the predicted value of the temperature of the process medium.

作为优选的技术方案，所述特征变量包括任意一个换热器内的工艺介质温度偏差值ei、工艺介质温度偏差变化率Δe_i、工艺介质实时温度检测值Tci以及循环冷却水系统中的冷却给水温度检测值Tgs、冷却给回水压差检测值P_Δj。As a preferred technical solution, the characteristic variables include the process medium temperature deviation ei in any heat exchanger, the process medium temperature deviation change rate Δe _i , the process medium real-time temperature detection value Tci, and the cooling feed water in the circulating cooling water system Temperature detection value Tgs, cooling water supply and return water pressure difference detection value P _Δj .

工艺介质温度偏差值等于对应所述工艺介质实时温度检测值与对应所述工艺介质温度设定值的差值；The temperature deviation value of the process medium is equal to the difference between the real-time temperature detection value corresponding to the process medium and the set value corresponding to the process medium temperature;

所述工艺介质温度偏差变化率为对应所述工艺介质相邻两个检测时段温度变化值与上一检测时段的比值。The temperature deviation change rate of the process medium corresponds to the ratio of the temperature change value of the process medium in two adjacent detection periods to the previous detection period.

在本实施例中，结合图8可以看出，步骤S72确定当前状态下的工艺介质最不利点的步骤为：In this embodiment, it can be seen from FIG. 8 that the step S72 determines the most unfavorable point of the process medium in the current state as follows:

S721：初始化，设M’个工艺介质温度偏差值组成一个差值小组，共计M’个工艺介质温度偏差值，令W_k＝M’；k＝1S721: Initialize, set M' process medium temperature deviation values to form a difference group, a total of M' process medium temperature deviation values, set W _k = M'; k = 1

S722：令W_k+1＝W_k+X，使W_k+1可以被M’整除，X为填充的差值无线大的空位；且X等于0～M’-1；S722: Let W _k+1 =W _k +X, so that W _k+1 can be divisible by M', and X is a gap filled with an infinitely large difference; and X is equal to 0 to M'-1;

S723：计算W_k+2＝W_k+1/M’S723: Calculate W _k+2 =W _k+1 /M'

S724：从W_k+2组中，采用交叉比较法，从每一组的M’个工艺介质温度偏差值中找出最小值，得到W_k+2个工艺介质温度偏差值；S724: From the W _k+2 group, use the cross-comparison method to find the minimum value from the M' process medium temperature deviation values in each group, and obtain W _k+2 process medium temperature deviation values;

S725：判断W_k+2是否等于1；若是，将该工艺介质温度偏差值作为工艺介质温度最不利点；否则，令k＝k+2；返回步骤S722。S725: Determine whether W _k+2 is equal to 1; if so, use the process medium temperature deviation value as the most unfavorable point of process medium temperature; otherwise, set k=k+2; return to step S722.

在本发明中，通过深层架构分别建立各换热器量测数据包括各换热器内工艺介质温度偏差及温度偏差变化率、冷却给水温度检测值以及冷却给回水压差设定值，与控制量预测量，即冷却给回水压差设定值之间的非线性映射关系。采用一种“预训练-参数微调”的两阶段离线训练学习方法，同时引入Dropout技术和RMSPROP技术对各换热器内工艺介质温度模型参数进行优化。训练后的模型能够依靠深层结构挖掘数据的隐藏模式，提取出有利于工艺介质温度预测控制效果的高阶特性。此外，该方法能够通过大量无标注样本的无监督训练提高模型泛化能力。In the present invention, the measurement data of each heat exchanger are respectively established through the deep structure, including the temperature deviation of the process medium in each heat exchanger and the temperature deviation change rate, the detection value of the cooling feed water temperature and the set value of the pressure difference of the cooling feed water and return water, and Pre-measurement of the control quantity, that is, the non-linear mapping relationship between the set values of the cooling water supply and return water pressure difference. A two-stage offline training and learning method of "pre-training-parameter fine-tuning" is adopted, and the Dropout technology and RMSPROP technology are introduced to optimize the temperature model parameters of the process medium in each heat exchanger. The trained model can rely on the deep structure to mine the hidden pattern of the data, and extract the high-order characteristics that are beneficial to the predictive control effect of the temperature of the process medium. In addition, the method can improve the generalization ability of the model through unsupervised training with a large number of unlabeled samples.

结合图6，所构建的SAE(stacked autoencoder，SAE)换热器内工艺介质训练好的深度学习网络的隐含层的层数以及每层隐含层的神经元个数对评估精度和离线训练时间有一定影响。以循环冷却水系统现场检测数据作为样本输入数据，逐层对隐含层神经元个数进行设置：首先确定第1层隐含层神经元的最优个数并固定，然后增加一层确定第2层隐含层神经元的最优个数，以此类推，直到平均百分比误差(MAPE)不再提高为止。Combined with Figure 6, the number of hidden layers of the deep learning network trained in the process medium in the constructed SAE (stacked autoencoder, SAE) heat exchanger and the number of neurons in each hidden layer have a significant impact on the evaluation accuracy and offline training Time plays a role. Taking the on-site detection data of the circulating cooling water system as the sample input data, the number of neurons in the hidden layer is set layer by layer: first determine the optimal number of neurons in the first layer of hidden layer and fix it, and then add a layer to determine the number of neurons in the first hidden layer. The optimal number of neurons in the hidden layer 2, and so on, until the average percentage error (MAPE) no longer increases.

为了使本发明的技术方案更加清楚,对本发明所用到的堆栈自动编码器的原理进行了解释。SAE通过逐层贪婪无监督预训练可以有效提取数据的高阶特征，更好地逼近复杂函数，并缩小参数寻优空间，能够快速得到网络参数，提高神经网络的深层特征学习能力。In order to make the technical solution of the present invention clearer, the principle of the stacked autoencoder used in the present invention is explained. SAE can effectively extract high-order features of data through layer-by-layer greedy unsupervised pre-training, better approximate complex functions, and reduce the parameter optimization space. It can quickly obtain network parameters and improve the deep feature learning ability of neural networks.

相比浅层机器学习算法，深层网络在处理高维数据时，因为其复杂的网络结构更加容易产生过拟合问题，从而限制模型的泛化能力。Dropout是一种主流的防过拟合技术，其基本思想为：在模型训练时随机选择一部分节点不工作，这些节点将保存上一次迭代的权值，并将输出置为0。这些被选择的节点在下次迭代的过程中又会恢复之前保留的权值，再次随机选择部分节点重复此过程。网络结构在每次迭代过程中都将发生一定的变化，采用Dropout技术，随机选取部分神经元暂时不工作，如图7所示，减少了特定节点之间的共同作用，减轻了网络输出对特定节点状态的依赖，从而防止过拟合。Compared with shallow machine learning algorithms, deep networks are more prone to overfitting problems due to their complex network structures when processing high-dimensional data, thus limiting the generalization ability of the model. Dropout is a mainstream anti-overfitting technology. Its basic idea is to randomly select some nodes to not work during model training. These nodes will save the weights of the previous iteration and set the output to 0. These selected nodes will restore the previously reserved weights in the next iteration, and then randomly select some nodes to repeat the process. The network structure will change to a certain extent during each iteration. Using Dropout technology, some neurons are randomly selected to temporarily stop working, as shown in Figure 7, which reduces the interaction between specific nodes and reduces the impact of network output on specific nodes. Dependence on node state, thus preventing overfitting.

应当指出的是，上述说明并非是对本发明的限制，本发明也并不仅限于上述举例，本技术领域的普通技术人员在本发明的实质范围内所做出的变化、改性、添加或替换，也应属于本发明的保护范围。It should be noted that the above description is not intended to limit the present invention, and the present invention is not limited to the above-mentioned examples. Those skilled in the art may make changes, modifications, additions or replacements within the scope of the present invention. It should also belong to the protection scope of the present invention.

1.一种循环冷却水系统工艺介质多温度目标设定值切换多参数预测控制算法，其中，循环冷却水系统包括沿着循环水路设置的N’个冷却塔、吸水池、给水泵机组、出水管组、M’个生产区换热器以及回水管组；在冷却塔内设置有冷却池；1. A multi-parameter predictive control algorithm for multi-temperature target setting value switching of process medium in a circulating cooling water system, wherein the circulating cooling water system includes N' cooling towers, suction pools, feed water pump units, outlet Water pipe group, M' production area heat exchangers and return water pipe group; a cooling pool is provided in the cooling tower; 所述出水管组包括N’根吸水池进水管、L根吸水池出水管、给水总管、M’根给水支管；The water outlet pipe group includes N' suction pool inlet pipes, L water suction pool outlet pipes, water supply main pipe, M' water supply branch pipes; 所述回水管组包括M’根回水支管、回水总管、N’根上塔回水管；The return pipe group includes M' backwater branch pipe, backwater main pipe, N' upper tower return pipe; 任一所述冷却塔经一根对应的吸水池进水管与所述吸水池连接，所述吸水池经L根并列的吸水池出水管与所述给水总管连接，所述给水总管经M’根给水支管向M’个生产区换热器一一对应供水；Any of the cooling towers is connected to the water-absorbing pool through a corresponding water-absorbing pool inlet pipe, and the water-absorbing pool is connected to the water supply main pipe through L parallel water-absorbing pool outlet pipes, and the water supply main pipe is connected to the water supply main pipe through M' The branch water supply pipe supplies water to the heat exchangers in M' production areas in one-to-one correspondence; 在所述吸水池出水管上并联设置有给水泵；A water feed pump is arranged in parallel on the water outlet pipe of the water absorption pool; 任一所述生产区换热器经一根对应的回水支管与所述回水总管连接，所述回水总管经并列的N’根上塔回水管与N’个冷却塔一一对应连接；Any one of the heat exchangers in the production area is connected to the return water main pipe through a corresponding return water branch pipe, and the return water main pipe is connected to N' cooling towers one by one through parallel N' upper tower return water pipes; M’个所述生产区换热器内均设置有工艺介质温度传感器，该工艺介质温度传感器用于获取各种工艺介质实时温度检测值；Process medium temperature sensors are all arranged in the heat exchangers of M' described production areas, and the process medium temperature sensors are used to obtain real-time temperature detection values of various process mediums; 在所述给水总管上设置有冷却给水温度传感器和冷却给水压力传感器，所述冷却给水温度传感器用于获取冷却给水温度检测值，所述冷却给水压力传感器用于获取冷却给水压力检测值；A cooling feed water temperature sensor and a cooling feed water pressure sensor are arranged on the feed water main pipe, the cooling feed water temperature sensor is used to obtain the detection value of the cooling feed water temperature, and the cooling feed water pressure sensor is used to obtain the detection value of the cooling feed water pressure; 在所述回水总管上设置有冷却回水温度传感器和冷却回水压力传感器，所述冷却回水温度传感器用于获取冷却回水温度检测值，所述冷却回水压力传感器用于获取冷却回水压力检测值；A cooling return water temperature sensor and a cooling return water pressure sensor are arranged on the return water main pipe, the cooling return water temperature sensor is used to obtain the detection value of the cooling return water temperature, and the cooling return water pressure sensor is used to obtain the cooling return water Water pressure detection value; 在N’根所述上塔回水管上分别设置有一个上塔阀；An upper tower valve is respectively arranged on the upper tower return pipes of the N' roots; 其特征在于：循环冷却水系统工艺介质多温度目标设定值切换多参数预测控制算法的步骤为：It is characterized in that: the step of multi-parameter predictive control algorithm switching of multi-temperature target setting value of process medium in circulating cooling water system is: S1：对M’个换热器进行编号，并获取循环冷却水系统在a个采样周期内运行产生的历史数据，并将获取的历史数据作为换热器内工艺介质温度预测模型的训练数据；S1: Number the M' heat exchangers, and obtain the historical data generated by the operation of the circulating cooling water system in a sampling period, and use the acquired historical data as the training data for the temperature prediction model of the process medium in the heat exchanger; S2：从历史数据中根据筛选条件筛选出特征变量，将换热器的历史数据中的工艺介质温度检测值、工艺介质温度偏差、工艺介质温度偏差变化率、冷却给水温度检测值、冷却给回水压差检测值作为输入数据，并进行数据归一化处理后得到归一数据集，并将该归一数据集划分为训练样本集和测试样本集；S2: Select the characteristic variables from the historical data according to the screening conditions, and collect the process medium temperature detection value, process medium temperature deviation, process medium temperature deviation change rate, cooling feed water temperature detection value, and cooling feed return in the historical data of the heat exchanger. The detected value of the water pressure difference is used as the input data, and the normalized data set is obtained after data normalization processing, and the normalized data set is divided into a training sample set and a test sample set; S3：基于堆叠自动编码器，对训练样本集进行逐层贪婪无监督预训练，得到深层神经网络初始化的输入层与隐含层的权值矩阵W、输入层与隐含层阈值矩阵B；S3: Based on the stacked autoencoder, perform layer-by-layer greedy unsupervised pre-training on the training sample set, and obtain the weight matrix W of the input layer and hidden layer initialized by the deep neural network, and the threshold matrix B of the input layer and hidden layer; S4：进行参数微调：对深层神经网络初始化的输入层与隐含层的权值矩阵W、输入层与隐含层阈值矩阵B进行微调，直到迭代次数达到迭代次数最大值为止，得到基于堆叠自动编码器的初始工艺介质温度预测模型；S4: Perform parameter fine-tuning: fine-tune the weight matrix W of the input layer and hidden layer initialized by the deep neural network, and the threshold matrix B of the input layer and hidden layer until the number of iterations reaches the maximum number of iterations, and the stack-based automatic The initial process medium temperature prediction model of the encoder; 参数微调时，采用自上而下的小批量RMSProp优化方法对深层神经网络初始化的输入层与隐含层的权值矩阵W、输入层与隐含层阈值矩阵B进行微调，具体步骤为：When fine-tuning parameters, the top-down small-batch RMSProp optimization method is used to fine-tune the weight matrix W of the input layer and hidden layer initialized by the deep neural network, and the threshold matrix B of the input layer and hidden layer. The specific steps are: S41，设置全局学习率l、衰减速率ρ，初始化累计变量r₁＝0，r₂＝0；S41, set global learning rate l, decay rate ρ, initialize cumulative variables r ₁ =0, r ₂ =0; S42，从训练集中选取包含m个样本的小批量数据集，根据误差损失函数，计算梯度：S42, select a small batch data set containing m samples from the training set, and calculate the gradient according to the error loss function:

S43，计算累计平方梯度，如式(9)所示：S43, calculate the cumulative square gradient, as shown in formula (9):

式中：⊙为逐元素乘积符号；In the formula: ⊙ is the element-wise product symbol; S44，分别更新权重和阈值参数：S44, updating weight and threshold parameters respectively:

S45，当迭代次数达到要求时，停止运算，否则返回第S42步继续执行计算；S45, when the number of iterations reaches the requirement, stop the operation, otherwise return to step S42 to continue the calculation; S5：使用测试样本数据集对步骤S4得到的初始工艺介质温度预测模型进行测试，得到基于堆叠自动编码器的工艺介质温度预测模型；S5: Use the test sample data set to test the initial process medium temperature prediction model obtained in step S4, and obtain a process medium temperature prediction model based on stacked autoencoders; S6：对步骤S5得到的工艺介质温度预测模型进行评估；S6: Evaluate the process medium temperature prediction model obtained in step S5; S7：获取循环冷却水系统当前状态下的当前数据，确定当前状态下的工艺介质最不利点，确定换热器控制对象，并结合工艺介质温度预测模型，得到对应的冷却给回水压差设定值和工艺介质温度预测值。S7: Obtain the current data of the circulating cooling water system in the current state, determine the most unfavorable point of the process medium in the current state, determine the control object of the heat exchanger, and combine the temperature prediction model of the process medium to obtain the corresponding cooling water supply and return water pressure difference setting Fixed value and predicted value of process medium temperature. 2.根据权利要求1所述的循环冷却水系统工艺介质多温度目标设定值切换多参数预测控制算法，其特征在于所述筛选条件是工艺介质温度处于安全且节能的温度值区间内的历史数据，所述安全且节能的温度值区间在工艺介质温度阈值区间内。2. The circulating cooling water system process medium multi-temperature target set value switching multi-parameter predictive control algorithm according to claim 1, characterized in that the screening condition is the history that the temperature of the process medium is in a safe and energy-saving temperature range data, the safe and energy-saving temperature range is within the process medium temperature threshold range. 3.根据权利要求1所述的循环冷却水系统工艺介质多温度目标设定值切换多参数预测控制算法，其特征在于步骤S3中，对训练样本集进行逐层贪婪无监督预训练时，将训练样本集分成P组小批量训练样本，依次进行训练，并采用Dropout技术，随机选取部分神经元暂停工作，依次迭代，逐层训练，得到深层神经网络初始化的输入层与隐含层的权值矩阵W、输入层与隐含层阈值矩阵B。3. The circulating cooling water system process medium multi-temperature target setting value switching multi-parameter predictive control algorithm according to claim 1, characterized in that in step S3, when the training sample set is carried out layer by layer greedy unsupervised pre-training, the The training sample set is divided into P groups of small-batch training samples, which are trained sequentially, and using Dropout technology, randomly select some neurons to suspend work, iterate in turn, and train layer by layer to obtain the weights of the input layer and hidden layer initialized by the deep neural network Matrix W, input layer and hidden layer threshold matrix B. 4.根据权利要求1所述的循环冷却水系统工艺介质多温度目标设定值切换多参数预测控制算法，其特征在于步骤S7的具体步骤为：4. The circulating cooling water system process medium multi-temperature target setting value switching multi-parameter predictive control algorithm according to claim 1, characterized in that the specific steps of step S7 are: S71：获取循环冷却水系统当前状态下的当前数据；S71: Obtain current data in the current state of the circulating cooling water system; S72：根据当前数据，确定当前状态下的工艺介质最不利点，S72: According to the current data, determine the most unfavorable point of the process medium under the current state, S73：确定该工艺介质最不利点对应的换热器j；S73: Determine the heat exchanger j corresponding to the most unfavorable point of the process medium; S74：获取的换热器j的特征变量；S74: The acquired characteristic variables of the heat exchanger j; S75：对换热器j的特征变量进行数据归一化处理；S75: Perform data normalization processing on the characteristic variables of the heat exchanger j; S76：结合工艺介质温度预测模型和步骤S75得到的数据，确定冷却给回水压差设定值和工艺介质温度预测值。S76: Combining the temperature prediction model of the process medium with the data obtained in step S75, determine the set value of the cooling feed-backwater pressure difference and the predicted value of the temperature of the process medium. 5.根据权利要求4所述的循环冷却水系统工艺介质多温度目标设定值切换多参数预测控制算法，其特征在于：所述特征变量包括任意一个换热器内的工艺介质温度偏差值、工艺介质温度偏差变化率、工艺介质实时温度检测值以及循环冷却水系统中的冷却给水温度检测值、冷却给回水压差检测值。5. The circulating cooling water system process medium multi-temperature target set value switching multi-parameter predictive control algorithm according to claim 4, characterized in that: the characteristic variable includes the temperature deviation value of the process medium in any one heat exchanger, The temperature deviation change rate of the process medium, the real-time temperature detection value of the process medium, the detection value of the cooling feed water temperature in the circulating cooling water system, and the detection value of the pressure difference of the cooling feed water and return water. 6.根据权利要求5所述的循环冷却水系统工艺介质多温度目标设定值切换多参数预测控制算法，其特征在于步骤S72确定当前状态下的工艺介质最不利点的步骤为：6. The circulating cooling water system process medium multi-temperature target setting value switching multi-parameter predictive control algorithm according to claim 5, characterized in that the step S72 determines the most unfavorable point of the process medium in the current state. S721：初始化，设M’个工艺介质温度偏差值组成一个差值小组，共计M’个工艺介质温度偏差值，令W_k＝M’；k＝1S721: Initialize, set M' process medium temperature deviation values to form a difference group, a total of M' process medium temperature deviation values, set W _k = M'; k = 1 S722：令W_k+1＝W_k+X，使W_k+1可以被M’整除，X为填充的差值无线大的空位；且X等于0～M’-1；S722: Let W _k+1 =W _k +X, so that W _k+1 can be divisible by M', and X is a gap filled with an infinitely large difference; and X is equal to 0 to M'-1; S723：计算W_k+2＝W_k+1/M’S723: Calculate W _k+2 =W _k+1 /M' S724：从W_k+2组中，采用交叉比较法，从每一组的M’个工艺介质温度偏差值中找出最小值，得到W_k+2个工艺介质温度偏差值；S724: From the W _k+2 group, use the cross-comparison method to find the minimum value from the M' process medium temperature deviation values in each group, and obtain W _k+2 process medium temperature deviation values; S725：判断W_k+2是否等于1；若是，将该工艺介质温度偏差值作为工艺介质温度最不利点；否则，令k＝k+2；返回步骤S722。S725: Determine whether W _k+2 is equal to 1; if so, use the process medium temperature deviation value as the most unfavorable point of process medium temperature; otherwise, set k=k+2; return to step S722.