CN-116306911-B - Distributed machine learning-based thermodynamic station load prediction and optimization control method

CN116306911BCN 116306911 BCN116306911 BCN 116306911BCN-116306911-B

Abstract

The invention discloses a distributed machine learning-based heating power station load prediction and optimization control method, which comprises the steps of setting corresponding edge computing devices near a plurality of heating power stations, constructing feature vectors based on heating power station load features and time sequence features, enabling a cloud server to initiate a global heat load prediction network model and send the feature vectors to each edge computing device, enabling the edge computing devices to conduct training of the heat load prediction network model by utilizing local data to obtain a local heat load prediction network model, enabling the cloud server to aggregate the local heat load prediction network models of the plurality of edge computing devices, updating the global heat load prediction network model and sending the local heat load prediction network model to each edge computing device, building a secondary network water supply temperature prediction model under the condition that total flow of a heating system and secondary network backwater temperature are unchanged, calculating a water supply temperature set value of a secondary network, and adjusting a primary network adjusting valve to control the water supply temperature of the secondary heat network.

Inventors

XIE JINFANG
JIN HEFENG
MU PEIHONG
ZHAO QIONG

Assignees

浙江英集动力科技有限公司

Dates

Publication Date: 20260505
Application Date: 20221215

Claims (7)

1. The utility model provides a heating power station load prediction and optimization control method based on distributed machine learning, which is characterized in that the heating power station load prediction and optimization control method includes: Step S1, setting corresponding edge computing devices near a plurality of heating stations, acquiring and preprocessing a heating station heat load influence data set to obtain a heating station load sequence, extracting additional features related to the time sequence, and constructing feature vectors based on heating station load features and time sequence features; s2, the cloud server utilizes the existing public data to initiate a global heat load prediction network model and broadcasts and transmits the model to each edge computing device in the federal learning system; S3, each edge computing device participating in training performs training of a thermal load prediction network model by using local data, obtains a local thermal load prediction network model, and uploads the local thermal load prediction network model to a cloud server; S4, the cloud server aggregates local heat load prediction network models of a plurality of edge computing devices, updates a global heat load prediction network model and transmits the global heat load prediction network model to each edge computing device; in the step S4, the cloud server aggregates local thermal load prediction network models of the plurality of edge computing devices, and updating the global thermal load prediction network model includes: the cloud server adopts FedAvg federal aggregation algorithm to conduct weighted aggregation on local heat load prediction network models of a plurality of edge computing devices, and the weight and threshold parameter updating of the global heat load prediction network model is completed; wherein, the FedAvg federal aggregation algorithm performs model weighted aggregation, expressed as: W t and W t+1 are respectively the weight and the activation threshold vector parameters of the global heat load prediction network model in the t and t+1 times of training, K is the number of heat stations, n is the total data sample amount of K heat stations, and n k is the data sample amount of heat station K; Predicting network model weight and activation threshold vector parameters for the local thermal load of the thermal station k for the t+1 round of training, eta is learning rate, g k is gradient of loss function of the local model of the thermal station k, L (-) is a loss function of the local thermal load prediction network model, X k is an input characteristic sequence of the local thermal load prediction network model of the heating power station k, y k is an actual thermal load value of the heating power station k, and w k is a weight and activation threshold vector parameter of the local thermal load prediction network model of the heating power station k; s5, repeating the steps S1-S4 until the model finally converges or reaches training times, and obtaining the load predicted value of each heating station through the global heat load predicted network model; Step S6, combining load predicted values of all heating stations, according to a heat balance equation, under the condition that the total flow of a heating system and the return water temperature of a secondary network are kept unchanged, adopting a distributed machine learning method to also establish a secondary network water supply temperature predicted model, calculating a water supply temperature set value of the secondary network, and adjusting a primary network regulating valve to control the water supply temperature of the secondary heating network, so that the actual water supply temperature value can follow the set temperature value to realize real-time control; In step S6, in combination with the load predicted values of the heat stations, according to a heat balance equation, under the condition of keeping the total flow of the heating system and the return water temperature of the secondary network unchanged, a distributed machine learning method is adopted to also establish a secondary network water supply temperature prediction model, and the calculation of the water supply temperature set value of the secondary network includes: The edge computing device is combined with the load predicted value of the heating power station, acquires and preprocesses a water supply temperature influence data set of the secondary network under the condition of keeping the total flow of the heating system and the return water temperature of the secondary network unchanged according to a heat balance equation, and then acquires a water supply temperature sequence, and extracts additional characteristics related to the time sequence, wherein the water supply temperature influence data set of the secondary network comprises the load predicted value of each heating power station, the return water temperature of the secondary network, the water supply temperature of the primary network, the return water temperature of the primary network, the flow rate of the primary network, the outdoor temperature, the outdoor wind speed and the outdoor humidity; the cloud server utilizes the existing public data to initiate a global secondary network water supply temperature prediction network model and broadcasts and transmits the model to each edge computing device in the federal learning system; Each edge computing device participating in training utilizes local data to train a secondary network water supply temperature prediction network model, obtains a local secondary network water supply temperature prediction network model, and uploads the local secondary network water supply temperature prediction network model to a cloud server; the cloud server aggregates the local secondary network water supply temperature prediction network models of the plurality of edge computing devices, updates the global secondary network water supply temperature prediction network model and transmits the global secondary network water supply temperature prediction network model to each edge computing device; repeating the above process until the model finally converges or reaches training times, and obtaining the secondary network water supply temperature set value corresponding to each heating power station through the global secondary network water supply temperature prediction network model.
2. The method for predicting and optimizing the load of a thermal power station according to claim 1, wherein in the step S1, the thermal power station load sequence is obtained after the thermal load influence data set of the thermal power station is acquired and preprocessed, the additional features related to the time sequence are extracted, and the feature vector is constructed based on the thermal power station load features and the time sequence features, and the method comprises the steps of: acquiring a heat load influence data set of the heating power station, wherein the heat load influence data set comprises historical water supply and return temperature, water supply and return flow, water supply and return pressure, heat load, historical outdoor temperature, outdoor wind speed and outdoor humidity of the heating power station; Carrying out wavelet packet decomposition and single reconstruction after carrying out missing data and outlier processing on a thermal load influence data set, and selecting wavelet basis functions and decomposition degrees to obtain decomposed low-frequency components and high-frequency component sequences; calculating importance of each feature in the load sequence of the heating power station by adopting XGBoost model, selecting the feature, and selecting the more important load feature as a load feature set of the heating power station; Considering month, hour, week and holiday at time t to form a time sequence feature set at time t; considering the load characteristic of the moment t of the whole point, the time distance from the moment t of the whole point and the load characteristic of the moment t of the next whole point, and the time distance from the moment t of the next whole point to form the load characteristic of the heating power station at the moment t; And forming a feature vector by the heat station load feature and the time sequence feature set at the moment t.
3. The heat station load prediction and optimization control method according to claim 1, wherein the heat load prediction network model adopts DeepAR autoregressive cyclic neural network, the interior of the heat load prediction network model comprises an LSTM model and a likelihood module, the likelihood module structure design comprises likelihood function selection, loss function determination and input layer neuron number setting, the likelihood function comprises a Gaussian likelihood function and a negative two-term likelihood function, and the loss function is defined as: z is a true value of the load of the heating station at a predicted moment, the load value of the heating station at any moment obeys Gaussian distribution, N-% (mu, sigma 2 ), mu is a mean value, sigma is a standard deviation, the number of neurons of an input layer is set according to the number of time feature sequences and the load value of the heating station at the last moment, the LSTM model comprises an input layer, an hidden layer, an output layer, an optimization algorithm, a loss function and an activation function, the number of neurons of the input layer is determined according to the sampling times of the load sequence of the heating station, the number of neurons of the hidden layer and the number of neurons of the hidden layer are determined by continuously adjusting parameters through a control variable method, the output layer neurons output the load value of the heating station at the next moment, the feature dimension is 1, the number of neurons of the output layer is 1, the optimization algorithm adopts an Adam optimization algorithm, the loss function adopts a difference between a mean square error measurement predicted value and the true value, the hidden layer activation function is set to be a sigmoid function, the output layer activation function is set to be a tanh function, the number of the self-regression circulation model is obtained by adopting a self-regression circulation model, and the self-circulation model is obtained through a parameter-regression algorithm in an ultra-circulation network model.
4. The method for predicting and optimizing load of heat station according to claim 1, wherein in step S3, training of the thermal load prediction network model is performed by using local data to obtain a local thermal load prediction network model, and the method comprises: The method comprises the steps of taking a feature vector comprising a heat station load feature and a time sequence feature as local data, inputting the local data into an LSTM model in a heat load prediction network model to train and obtain an output h i,t ＝h(h i,t-1 ,z i,t-1 ,x i,t theta, wherein h is an implicit layer function, the inside of the thermal load prediction network model is realized by adopting a multi-layer cyclic neural network, the implicit layer function is parameterized through a parameter theta, in the training process, at each time step t, h i,t-1 is the output state of the last time step of the heat load prediction network model, z i,t-1 is the value of the last time step of a measured starting point and represents the actually observed data value at the last moment, and x i,t is the value of the heat station load sequence i on the time step t and represents the network input state; the gaussian distribution parameters μ and σ of the time series over the future time steps are calculated as: And B μ and b σ are all-connection layer bias vectors; And determining Gaussian distribution according to the Gaussian distribution parameters mu and sigma, and solving the binary number of the Gaussian distribution as a predicted value z i,t of the thermal station heat load sequence at the time t.
5. The heat station load prediction and optimization control method according to claim 1, further comprising: When a new heating station is added, the movable thermal load prediction network model and the secondary network water supply temperature prediction network model are trained by selecting similar heating stations from the existing heating stations and utilizing the data of the similar heating stations, and then the thermal load prediction network model and the secondary network water supply temperature prediction network model are further finely adjusted based on the operation data of the new heating stations.
6. The method for predicting and optimizing load of a thermal station according to claim 5, wherein the selecting similar thermal stations considers the service objects, the positions and functions of the thermal stations, and the areas of the heating medium and the thermal stations for connecting the heat supply network; Training a movable thermal load prediction network model and a secondary network water supply temperature prediction network model by using data of similar thermal stations, namely training a global thermal load prediction network model and a secondary network water supply temperature prediction network model by using data of the similar thermal stations, and migrating weight and activation threshold vector parameters of the trained models into edge computing devices corresponding to the new thermal stations; the method comprises the steps of further fine-tuning a thermal load prediction network model and a secondary network water supply temperature prediction network model based on operation data of a new thermal station, wherein the weight and the activation threshold vector parameters of the migrated model are used as initial parameters for training a local federal model, then, the new gradient of the model is calculated by using the local operation data of the new thermal station, and the model parameters are adjusted to obtain the thermal load prediction network model and the secondary network water supply temperature prediction network model of the new thermal station.
7. The method for predicting and optimizing load of a heat station according to claim 1, wherein the cloud server and the edge computing device encrypt data and transmit the encrypted data to a receiving party, and the receiving party decrypts the encrypted data to obtain plaintext data when transmitting the data, and the cloud server and the edge computing device are provided with a password unit, wherein the password unit is preset with encryption and decryption algorithms, and the password unit at least comprises a symmetric encryption algorithm and an asymmetric encryption algorithm.

Description

Distributed machine learning-based thermodynamic station load prediction and optimization control method Technical Field The invention belongs to the technical field of intelligent heat supply, and particularly relates to a thermodynamic station load prediction and optimization control method based on distributed machine learning. Background With the proposal of the aim of constructing an economic society in China, central heating gradually becomes a main heating mode in winter in the north. The central heating system has the characteristics of complex structure, nonlinearity, large hysteresis, large inertia, time variability, uncertainty and the like, and is very necessary to realize the on-demand heating of a heating network, improve the energy utilization rate, optimize the heating regulation strategy and the accurate heat load prediction. Therefore, based on the characteristics of the heating system, the heat load prediction and the optimization control become particularly important. The traditional method takes heat load historical data of a heat exchange station in a heat supply system as data support, predicts the heat load by machine learning, and provides theoretical basis for the operation regulation of a later system, thereby improving the control quality of a heat supply network. The machine learning method has strong fitting capability, can fully reflect the nonlinear characteristic of the heat load, not only provides reliable heat load data for a control system, but also can effectively support heat load adjustment. However, collecting, storing, exchanging real-time operational data of a large number of thermal stations presents a serious risk of privacy disclosure. In addition, how to accurately and rapidly predict the heat load is not only crucial to the stable and normal operation of the heat supply system, but also can have great influence on the production and life of the whole society. Therefore, how to solve the problem of data leakage, and at the same time, ensure the accuracy of the prediction model and improve the training speed of the model is an urgent problem to be solved at present. Based on the technical problems, a new thermodynamic station load prediction and optimization control method based on distributed machine learning needs to be designed. Disclosure of Invention The invention aims to solve the technical problems of overcoming the defects of the prior art, and provides a distributed machine learning-based thermodynamic station load prediction and optimization control method, which can combine federal learning and edge cloud cooperative technology to build a prediction model, perform local training and federal aggregation processes of the prediction model, and model the time series data by using additional features of the time series data, so that the time series data are predicted more accurately, feature vectors are built based on the thermodynamic station load features and the time series features, the prediction model is trained, and the model prediction precision is improved. In order to solve the technical problems, the technical scheme of the invention is as follows: the invention provides a heat station load prediction and optimization control method based on distributed machine learning, which comprises the following steps: Step S1, setting corresponding edge computing devices near a plurality of heating stations, acquiring and preprocessing a heating station heat load influence data set to obtain a heating station load sequence, extracting additional features related to the time sequence, and constructing feature vectors based on heating station load features and time sequence features; s2, the cloud server utilizes the existing public data to initiate a global heat load prediction network model and broadcasts and transmits the model to each edge computing device in the federal learning system; S3, each edge computing device participating in training performs training of a thermal load prediction network model by using local data, obtains a local thermal load prediction network model, and uploads the local thermal load prediction network model to a cloud server; S4, the cloud server aggregates local heat load prediction network models of a plurality of edge computing devices, updates a global heat load prediction network model and transmits the global heat load prediction network model to each edge computing device; s5, repeating the steps S1-S4 until the model finally converges or reaches training times, and obtaining the load predicted value of each heating station through the global heat load predicted network model; and S6, combining load predicted values of all heating stations, according to a heat balance equation, under the condition that the total flow of the heating system and the return water temperature of the secondary network are kept unchanged, adopting a distributed machine learning method to also establish a secondary network water supply tempera