Comprehensive energy system multi-element load short-term prediction method based on multi-task learning

Abstract

The invention relates to an integrated energy system multi-element load short-term prediction method based on multi-task learning, and the method comprises the steps: constructing a prediction model of a bidirectional long-short-term memory network based on multi-task learning, training the prediction model through preprocessed sample data, updating the weight and bias of each network layer in the prediction model, and carrying out the prediction of the multi-element load of the integrated energy system. And performing multi-element load short-term prediction on the integrated energy system by using the trained prediction model. According to the method, prior complex mechanism knowledge is not needed, and higher accuracy is achieved; on the basis of high prediction precision, the prediction performance of balancing subtasks is guaranteed, the coupling between loads is fully mined, and the prediction performance is improved; the model can automatically learn characteristics of load fluctuation; and accurate prediction of the short-term load demand of the multi-element load of the integrated energy system is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of calculation, extrapolation or counting, and in particular to a short-term prediction method for multi-element loads of an integrated energy system based on multi-task learning. Background Art

[0002] The integrated energy system is the development trend of future energy consumption. It is a way to effectively improve energy utilization and reduce carbon emissions through the coordinated planning and dispatch of multiple energy sources such as electricity, cooling and heat. At a time when low-carbon, environmentally friendly and sustainable development has become a top priority, the integrated energy system is of great significance for achieving carbon peak and carbon neutrality as soon as possible, promoting ecological civilization construction and energy revolution and low-carbon economic development, ensuring rural construction, and responding to global climate change.

[0003] In the field of coordinated planning and scheduling of integrated energy systems, short-term prediction of multiple loads is an extremely important link, and its accuracy directly affects the coordinated planning and scheduling effect of integrated energy systems. However, in the prediction of multiple loads, the load demand on the load side is affected by factors such as user personal behavior, weather, and holidays, which leads to large fluctuations in load demand. Considering the coupling relationship between different loads, rapid response to load fluctuations and exploitation of the coupling between loads can improve the accuracy of load prediction. Therefore, in the prediction of multiple loads in integrated energy systems, rapid response to load fluctuations and exploitation of the coupling between loads have very important theoretical and practical significance.

[0004] There is no mature method for mining the coupling relationship between multiple loads in the integrated energy system. The existing technology generally adopts the method of correlation analysis of load data, but this method is difficult to mine the complex coupling relationship between loads. Furthermore, the existing technology generally adopts the method of quickly responding to load fluctuations by releasing electric energy through energy storage systems when loads fluctuate, but this method requires the construction of a huge energy storage system, which is costly.

[0005] The above situations restrict the accuracy of short-term forecasting of multiple loads in the integrated energy system, and thus fail to realize the value of the forecast. Summary of the invention

[0006] The present invention solves the problems existing in the prior art and provides a short-term forecasting method for multiple loads in an integrated energy system based on multi-task learning that combines load coupling analysis and load sudden change detection. Multi-task learning is used to explore the coupling between loads to improve the performance of the model, and load sudden change analysis is integrated to quickly respond to load fluctuations.

[0007] The technical solution adopted by the present invention is a short-term prediction method for multiple loads in an integrated energy system based on multi-task learning, which constructs a prediction model of a bidirectional long short-term memory network based on multi-task learning, trains the prediction model with preprocessed sample data, updates the weights and biases of each network layer in the prediction model, and uses the trained prediction model to perform short-term prediction of multiple loads in an integrated energy system.

[0008] Preferably, in preprocessing the sample data, the load data and influencing factor data of the multi-element load of the integrated energy system are collected, and abnormal data processing is performed, and input variables are selected based on the maximum mutual information coefficient method for the data after abnormal data processing, and dimensionality reduction processing is performed on the data based on the selected input variables;

[0009] The prediction model is used to obtain the coupling correlation between the load data after dimensionality reduction processing.

[0010] Preferably, the prediction model comprises an input layer, a coupling analysis-multi-granularity feature extraction layer, a gating layer and an output layer which are arranged in sequence; the coupling analysis-multi-granularity feature extraction layer is arranged based on a bidirectional long short-term memory network.

[0011] Preferably, the coupling analysis-multi-granularity feature extraction layer includes a juxtaposed shared layer and one or more expert layers, and any of the expert layers matches the type of sample data; a corresponding gating unit is provided in the gating layer cooperating with any of the expert layers, and the shared layer is arranged in coordination with all the gating units.

[0012] Preferably, the number of types of load data corresponds to the number of expert layers.

[0013] Preferably, the load data includes electric load data, cooling load data and heating load data, and the coupling associations among the load data are pairwise couplings among the electric load data, cooling load data and heating load data.

[0014] Preferably, the multi-task loss function of the prediction model is positively correlated with the electric load prediction subtask loss function, the cooling load prediction subtask loss function and the heating load prediction subtask loss function.

[0015] Preferably, the load mutation sensitivity index and the standard deviation within the rolling window segment are used to judge the load change state of the integrated energy system, and the prediction model is trained with a corresponding learning rate in accordance with the load change state.

[0016] Preferably, the load mutation sensitivity index satisfies:

[0017]

[0018] Among them, t 1 is the time point when the load mutation occurs, t 2To determine the time point when the error is less than the preset value, if the preset value is set to 0.05, E after The mean square error before the load mutation occurs, E before The mean square error after a sudden load change occurs.

[0019] Preferably, the data after dimension reduction processing includes local temperature data, local air pressure data, local humidity data and local dew point temperature data.

[0020] The present invention relates to a method for short-term prediction of multiple loads in an integrated energy system based on multi-task learning. A prediction model of a bidirectional long short-term memory network based on multi-task learning is constructed, the prediction model is trained with preprocessed sample data, the weights and biases of each network layer in the prediction model are updated, and the short-term prediction of multiple loads in the integrated energy system is performed with the trained prediction model.

[0021] The beneficial effects of the present invention are:

[0022] (1) Compared with the traditional mechanism-based load forecasting model, it does not require prior complex mechanism knowledge and has higher accuracy;

[0023] (2) A multi-task learning algorithm with progressive hierarchical extraction is used to exchange information and coordinate optimization through the shared layer, so that the model can discover and utilize the correlation and shared information between tasks, and achieve the purpose of mining the coupling between load data. The expert layer is used to process different loads in a more detailed manner, so that the model can capture and utilize specific information or patterns related to each task, thereby achieving the purpose of balancing subtasks, and then ensuring the prediction performance of balancing subtasks on the basis of high prediction accuracy, fully mining the coupling between loads, and improving prediction performance;

[0024] (3) A load sudden change analysis method is proposed to determine whether the load has suddenly changed, and to quickly respond to load sudden changes and handle load fluctuations. By using different learning rates for different load fluctuation stages, the model can automatically learn the characteristics of load fluctuations.

[0025] (4) Accurately predict the short-term load demand of multiple loads in the integrated energy system. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a prediction flow chart of the present invention;

[0027] Figure 2 It is a structural diagram of the comprehensive energy system in the present invention;

[0028] Figure 3 This is a structural diagram of the multi-task learning model in the present invention. DETAILED DESCRIPTION

[0029] The present invention is further described in detail below in conjunction with embodiments, but the protection scope of the present invention is not limited thereto.

[0030] The present invention relates to a method for short-term prediction of multiple loads in an integrated energy system based on multi-task learning. A prediction model of a bidirectional long short-term memory network based on multi-task learning is constructed, the prediction model is trained with preprocessed sample data, the weights and biases of each network layer in the prediction model are updated, and the short-term prediction of multiple loads in the integrated energy system is performed with the trained prediction model.

[0031] In the present invention, a short-term forecasting method for multiple loads of an integrated energy system based on multi-task learning with progressive hierarchical extraction is used to analyze the coupling between different loads and balance the forecasting performance of subtasks. On this basis, steep load analysis is introduced to meet the demand for rapid response to load fluctuations.

[0032] The present invention comprises the following steps in practical application:

[0033] (1) Preprocess sample data;

[0034] (2) Construct a prediction model based on a bidirectional long short-term memory network with multi-task learning;

[0035] (3) Input the preprocessed sample data into the prediction model for training;

[0036] (4) Perform load forecasting using the trained forecasting model.

[0037] In actual application, there is no order in which the above steps (1) and (2) are performed.

[0038] The steps are described in detail below with reference to the embodiments.

[0039] (1) Preprocess sample data;

[0040] In the present invention, sample data refers to input and output data pairs selected according to given rules, and the given rules include:

[0041] a) The selected sample data should reach a certain scale. Too little data will affect the fitting accuracy of the prediction model; too much data will lead to a longer training time;

[0042] b) The selected samples should include data under various operating conditions of the integrated energy system, cover the known fluctuation range as much as possible, and exclude extreme abnormal data.

[0043] Preprocessing the sample data includes the following steps:

[0044] (1-1) Collect load data and influencing factor data of multiple loads in the integrated energy system;

[0045] The load data here include electric load data, cooling load data and heating load data; of course, in practical applications, those skilled in the art are capable of splitting the load data into more categories, and correspondingly, the number of expert layers needs to be adjusted.

[0046] Influencing factor data refers to data that may affect load changes, including but not limited to local temperature data, air pressure data, humidity data, dew point temperature data, near-ground wind speed data, wind direction data, cloud cover data, solar radiation data, holiday data, etc.;

[0047] (1-2) Process abnormal data;

[0048] Z-Score is used to denoise the influencing factor data and load data and to select outliers. The outliers are replaced by the mean of the four normal values ​​before and after them, and then the data is normalized to avoid the model being too complicated, reducing the accuracy and increasing the calculation time caused by abnormal data.

[0049] (1-3) For the data after abnormal data processing, input variables are selected based on the maximum mutual information coefficient method, and dimensionality reduction is performed on the data based on the selected input variables;

[0050] The historical operation data of 12 input variables that have been roughly screened and processed are collected. Based on these data, the data space is gridded, the mutual information value of each grid data is calculated and normalized to obtain the maximum mutual information coefficient (MIC).

[0051]

[0052]

[0053] Among them, I(X; Y) is the mutual information between variables X and Y, p(x; y) is the joint probability density between variables, MIC[X; Y] is the maximum information coefficient between variables, and B is the number of grids, which is generally taken as 0.6 of the total amount of data;

[0054] Considering that in the actual system operation process, when the MIC value is greater than 0.4, it can be considered that there is a significant correlation between the two variables, so the influencing factor data with MIC values ​​greater than 0.4 for the three loads are retained;

[0055] The data after the dimension reduction process includes local temperature data, local air pressure data, local humidity data and local dew point temperature data; these variables are used as input variables of the subsequent model.

[0056] The prediction model is used to obtain the coupling association between the load data after the dimension reduction process, that is, the pairwise coupling between the electric load data, the cooling load data and the heating load data;

[0057] Specifically, for different load data, the load demand is not only affected by other non-influencing factors such as weather data, but there is also mutual influence between loads. For example, in daily life, heating by electricity or heating corresponds to the electric-thermal coupling relationship, so it is necessary to explore the coupling between loads.

[0058] (2) Construct a prediction model based on a bidirectional long short-term memory network with multi-task learning;

[0059] The prediction model includes an input layer, a coupling analysis-multi-granularity feature extraction layer, a gating layer and an output layer which are arranged in sequence; the coupling analysis-multi-granularity feature extraction layer is arranged based on a bidirectional long short-term memory network.

[0060] The coupling analysis-multi-granularity feature extraction layer includes a juxtaposed shared layer and one or more expert layers, any of which matches the type of sample data; a corresponding gating unit is provided in the gating layer that cooperates with any of the expert layers, and the shared layer is arranged in coordination with all the gating units.

[0061] In the present invention, the input layer is used to receive processed input data and convert the input data into a supervised learning format;

[0062] In the coupling analysis-multi-granularity feature extraction layer, the shared layer is used to mine the coupling between loads and capture the overall temporal characteristics of the input data. It consists of a bidirectional long short-term memory network with 128 neurons.

[0063] In the coupling analysis-multi-granularity feature extraction layer, the expert layer is used to extract more refined and specific features for specific tasks, and is used to optimize the needs of a specific task. It is also composed of a bidirectional long short-term memory network. Taking the three loads predicted as three subtasks as an example, there are three expert layers here, corresponding to the types of sample data, namely electric load data, cooling load data and heat load data; among them, because the cold and heat loads have similar trends in load demand and fluctuation, the number of experts set for the corresponding expert layers of electric load, cooling load and heat load are 6, 3, and 3 respectively, and the number of neurons in each expert layer is set to 32; by setting up a separate expert layer, the performance of subtasks can be balanced while ensuring the prediction accuracy; the data output by the input layer is input to the expert layer and the shared layer at the same time;

[0064] The coupling analysis here - the multi-granularity feature extraction layer is a progressive hierarchical extraction method. The multi-task learning model not only has a parameter sharing layer, but also has a corresponding expert task layer for each subtask. This makes it possible to share parameters and learn the coupling between load data through the sharing layer in the short-term prediction of multi-loads in the integrated energy system, and each unique expert task layer can learn different data features of different loads, so as to make more detailed predictions of different loads.

[0065] The gating layer is used to decide how to weight and adjust the data input to the gating layer. It is constructed by multiple fully connected layers, where the output of the gating layer is as follows:

[0066]

[0067] Among them, σ(·) activation function is the sigmoid function, W is the weight matrix of the gating layer, b is the bias of the gating layer, and x load (t) is the output of the shared layer, x i (t) is the output of the expert layer;

[0068] The output layer is the prediction layer, which is used to obtain and output the final prediction result. It is a fully connected layer.

[0069] (3) Input the preprocessed sample data into the prediction model for training;

[0070] The multi-task loss function of the prediction model is positively correlated with the electric load prediction subtask loss function, the cooling load prediction subtask loss function and the heating load prediction subtask loss function.

[0071] The loss functions of the three subtasks are as follows:

[0072]

[0073]

[0074] in, Respectively represent the predicted values ​​of electric load, cooling load and heating load, x 1 , x 2 , x 3 They represent the true values ​​of electric load, cooling load and heating load respectively, i is the i-th data of each load, and N is the total number of samples;

[0075] The total task loss function satisfies,

[0076] l=α 1 L 1 +α 2 L 2 +α 3 L 3

[0077] Among them, α 1 , α 2 , α 3 According to human experience, they are set to 0.4, 0.2, and 0.4 respectively;

[0078] Through model training, we find the model parameters that minimize the total task loss function.

[0079] In the training process, the learning rate of the Adam optimizer is initially set to 0.001, and other parameters are all default values. Furthermore, a load sudden change analysis is introduced, and different learning rates are applied to respond to load fluctuations at different stages of load change. Specifically, a load real-time sensitivity index LSSI and the standard deviation in the rolling window segment are introduced. The calculation of LSSI is based on the change of the mean square error, and the length of the rolling window is based on the fluctuation time range of the load data, such as being set to 16, and the thresholds of LSSI and the standard deviation are set. In the process of model training, when the value of LSSI is greater than the threshold and the standard deviation in the rolling window segment is also greater than the threshold, it is determined that the load fluctuates significantly, thereby adjusting the model learning rate to 0.0001. When the conditions are not met, the model learning rate is adjusted to 0.001, that is, the learning rate of the model is 0.001 under normal circumstances.

[0080] The load change state of the integrated energy system is judged by the load mutation sensitivity index and the standard deviation within the rolling window segment, and the prediction model is trained with the corresponding learning rate in accordance with the load change state.

[0081] The load mutation sensitivity index satisfies:

[0082]

[0083] Among them, t 1 is the time point when the load mutation occurs, t 2 To determine the time point when the error is less than the preset value, that is, when the error tends to be stable, E after The mean square error before the load mutation occurs, E before The mean square error after a load mutation occurs; the larger the load mutation sensitivity index is, the faster the model can respond to load fluctuations.

[0084] By changing the model learning rate, the model can perform differently in different load change stages, which can make the model stable when the load changes drastically and accelerate convergence when the load is stable, effectively improving the accuracy and stability of load forecasting;

[0085] Furthermore, the prediction model is trained using the gradient descent algorithm, the model parameters are updated, the gradient of the loss function with respect to the parameters is calculated, and the weight matrix and bias term W,b of the prediction model are updated as follows,

[0086]

[0087]

[0088] Where η is the learning rate of the gradient descent algorithm, W(k) is the weight matrix corresponding to the sample data at the kth moment, W(k-1) is the weight matrix corresponding to the sample data at the k-1th moment, b(k) is the bias term corresponding to the sample data at the kth moment, and b(k-1) is the bias term corresponding to the sample data at the k-1th moment. represents the partial differential symbol, and L is the total task loss function.

[0089] (4) Perform load forecasting using the trained forecasting model.

[0090] In order to implement the above-mentioned embodiment, a computer-readable storage medium is proposed, on which a short-term prediction program for multiple loads of an integrated energy system based on multi-task learning is stored. When the short-term prediction program for multiple loads of an integrated energy system based on multi-task learning is executed by a processor, the short-term prediction method for multiple loads of an integrated energy system based on multi-task learning is implemented.

[0091] In order to implement the above-mentioned embodiments, a computer device is also proposed, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the short-term prediction method for multiple loads of an integrated energy system based on multi-task learning as described above is implemented.

[0092] In practical application, the multi-task learning-based short-term forecasting method for integrated energy system multi-load includes the following processes:

[0093] Parameter initialization: Set appropriate initial values ​​for the hyperparameters in the proposed method, import the reduced-dimensional integrated energy system multivariate load data and sample data of influencing factors in the import interface of the control computer model, and divide the model into training set, validation set and test set; and set the prediction duration of the future H hours;

[0094] Offline training: First, the algorithm is written in the form of code and input into the control computer to preliminarily build the model framework. Then, the imported multi-dimensional load data of the integrated energy system and the sample data of the influencing factors after dimensionality reduction are input into the algorithm to train the algorithm to obtain the final appropriate hyperparameter values. Finally, the trained algorithm model can be obtained.

[0095] Online prediction: Start the GPU of the control computer, read the trained model network parameters and the set value of the future prediction time H. By measuring the variable data of the integrated energy system online and executing the corresponding program of the algorithm, it is possible to predict in real time whether the various energy sources of the integrated energy system meet the demand.

[0096] Through the short-term forecasting method, medium, equipment and application of multiple loads in integrated energy system, the problems of low prediction accuracy, difficulty in exploring coupling between loads and high economic cost in the existing short-term forecasting of multiple loads in integrated energy system can be overcome, which is easy to operate and explainable.

[0097] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A short-term forecasting method for multi-element loads in an integrated energy system based on multi-task learning, characterized by: A prediction model of a bidirectional long short-term memory network based on multi-task learning is constructed, the prediction model is trained with preprocessed sample data, the weights and biases of each network layer in the prediction model are updated, and the trained prediction model is used to perform short-term prediction of multivariate loads in an integrated energy system.

2. According to claim 1, a method for short-term forecasting of multiple loads in an integrated energy system based on multi-task learning is characterized by: In preprocessing the sample data, the load data and influencing factor data of the multi-load of the integrated energy system are collected, and abnormal data processing is performed, and input variables are selected based on the maximum mutual information coefficient method for the data after abnormal data processing, and dimensionality reduction processing is performed on the data based on the selected input variables; The prediction model is used to obtain the coupling correlation between the load data after dimensionality reduction processing.

3. The method for short-term forecasting of multiple loads in an integrated energy system based on multi-task learning according to claim 1 is characterized in that: The prediction model includes an input layer, a coupling analysis-multi-granularity feature extraction layer, a gating layer and an output layer which are arranged in sequence; the coupling analysis-multi-granularity feature extraction layer is arranged based on a bidirectional long short-term memory network.

4. The method for short-term forecasting of multiple loads in an integrated energy system based on multi-task learning according to claim 3 is characterized by: The coupling analysis-multi-granularity feature extraction layer includes a juxtaposed shared layer and one or more expert layers, any of which matches the type of sample data; a corresponding gating unit is provided in the gating layer that cooperates with any of the expert layers, and the shared layer is arranged in coordination with all the gating units.

5. The method for short-term forecasting of multiple loads in an integrated energy system based on multi-task learning according to claim 4 is characterized in that: The number of types of the load data corresponds to the number of expert layers.

6. The method for short-term forecasting of multiple loads in an integrated energy system based on multi-task learning according to claim 5 is characterized by: The load data includes electric load data, cooling load data and heating load data, and the coupling association between the load data is the pairwise coupling between the electric load data, cooling load data and heating load data.

7. The method for short-term forecasting of multiple loads in an integrated energy system based on multi-task learning according to claim 6 is characterized by: The multi-task loss function of the prediction model is positively correlated with the electric load prediction subtask loss function, the cooling load prediction subtask loss function and the heating load prediction subtask loss function.

8. The method for short-term forecasting of multiple loads in an integrated energy system based on multi-task learning according to claim 1 is characterized by: The load change state of the integrated energy system is judged by the load mutation sensitivity index and the standard deviation within the rolling window segment, and the prediction model is trained with the corresponding learning rate in accordance with the load change state.

9. The method for short-term forecasting of multiple loads in an integrated energy system based on multi-task learning according to claim 8 is characterized by: The load mutation sensitivity index satisfies: Among them, t1 is the time point when the load mutation occurs, t2 is the time point when the error is determined to be less than the preset value, and E after The mean square error before the load mutation occurs, E before The mean square error after a sudden load change occurs.

10. The method for short-term forecasting of multiple loads in an integrated energy system based on multi-task learning according to claim 2, characterized in that: The data after the dimension reduction processing includes local temperature data, local air pressure data, local humidity data and local dew point temperature data.

Images