Multi-task learning based comprehensive energy system multi-element load short-term prediction method

Abstract

The present application relates to a comprehensive energy system multi-element load short-term prediction method 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 by using preprocessed sample data, the weights and biases of each network layer in the prediction model are updated, and the trained prediction model is used for short-term prediction of multi-element load of a comprehensive energy system. The present application does not require prior complex mechanism knowledge, and has higher accuracy; on the basis of high prediction accuracy, the prediction performance of the balanced sub-task is ensured, the coupling between loads is fully tapped, and the prediction performance is improved; the model can automatically learn the characteristics of load fluctuation; and accurate short-term load demand prediction of multi-element load of a comprehensive energy system is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of calculation, estimation or counting, in particular to a multi-task learning-based comprehensive energy system multi-element load short-term prediction method. BACKGROUND

[0002] The comprehensive energy system is a development trend of future energy consumption, which is a way to effectively improve energy utilization and reduce carbon emissions through the coordinated planning and scheduling of electricity, cold, heat and other multiple energies. In the current situation where low-carbon, environmental protection and sustainable development have become a top priority, the comprehensive energy system is of great significance to achieving carbon peak and carbon neutrality as soon as possible, promoting ecological civilization construction, 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 comprehensive energy systems, multi-element load short-term prediction is an extremely important link, and its accuracy directly affects the coordinated planning and scheduling effect of comprehensive energy systems. However, in multi-element load prediction, because the load demand on the load side is affected by user behavior, weather, holidays and other factors, the load demand fluctuates greatly, and considering the coupling relationship between different loads, quickly responding to load fluctuations and mining the coupling between loads can improve the load prediction accuracy. Therefore, in the multi-element load prediction of comprehensive energy systems, quickly responding to load fluctuations and mining the coupling between loads have very important theoretical and practical significance.

[0004] There is no mature method for how to mine the coupling relationship between multi-element loads of comprehensive energy systems in the prior art, and the existing technology generally uses correlation analysis on load data, but this method is difficult to mine the complex coupling relationship between loads. Further, the existing technology generally uses the electric energy released by the energy storage system during load fluctuation to quickly respond to load fluctuations, but this method requires the construction of a large energy storage system, which is costly.

[0005] The above situations all restrict the accuracy of multi-element load short-term prediction of comprehensive energy systems, and thus the value of prediction cannot be realized. SUMMARY

[0006] The present application solves the problems in the prior art and provides a multi-task learning-based comprehensive energy system multi-element load short-term prediction method combining load coupling analysis and load steep change detection, which uses multi-task learning to mine the coupling between loads, improves the performance of the model, and integrates load steep change analysis to quickly respond to load fluctuations.

[0007] The technical scheme adopted by the present application is a comprehensive energy system multi-element load short-term prediction method 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 by using preprocessed sample data, the weights and biases of each network layer in the prediction model are updated, and the trained prediction model is used for short-term prediction of the multi-element load of the comprehensive energy system.

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

[0009] The coupling relationship between the load data after dimension reduction is obtained by the prediction model.

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

[0011] Preferably, the coupling analysis-multigranularity feature extraction layer comprises a shared layer and one or more expert layers arranged in parallel, any expert layer matches the type of sample data; the gate layer matched with any expert layer is provided with a corresponding gate unit, and the shared layer is arranged in cooperation with all gate units.

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

[0013] Preferably, the load data comprises electric load data, cold load data and heat load data, and the coupling relationship between the load data is the two-by-two coupling between the electric load data, the cold load data and the heat load data.

[0014] Preferably, the multi-task loss function of the prediction model is positively correlated with an electric load prediction sub-task loss function, a cold load prediction sub-task loss function and a heat load prediction sub-task loss function.

[0015] Preferably, the load change state of the comprehensive energy system is judged by using a load mutation sensitivity index and a standard deviation in a rolling window segment, and the prediction model is trained by using a corresponding learning rate matched with the load change state.

[0016] Preferably, the load mutation sensitivity index satisfies

[0017]

[0018] Wherein, t1 is the time point of load mutation, t2 is the time point of determining that the error is less than the preset value, such as the preset value is 0.05, E after The mean square error before the load mutation occurs before The mean square error after the load mutation occurs.

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

[0020] The present application relates to a kind of based on multi-task learning's comprehensive energy system multi-element load short-term prediction method, constructs the prediction model of bidirectional long short-term memory network based on multi-task learning, with preprocessed sample data training the prediction model, update the weight and bias of each network layer in prediction model, with the trained prediction model carries out comprehensive energy system multi-element load short-term prediction.

[0021] The beneficial effects of the present application are that:

[0022] (1) compared with the traditional load prediction model based on mechanism, no prior complex mechanism knowledge is needed, and it has higher accuracy;

[0023] (2) the multi-task learning algorithm of progressive hierarchical extraction is adopted, information exchange and collaborative optimization are carried out through the shared layer, so that the model discovers and utilizes the correlation and shared information between tasks, achieves the purpose of mining the coupling between load data, uses expert layer to process different loads more finely, so that the model can capture and utilize specific information or patterns related to each task, so as to realize the purpose of balanced subtask, and then guarantee the prediction performance of balanced subtask on the basis of high prediction accuracy, fully mine the coupling between loads, and improve the prediction performance;

[0024] (3) load steep change is proposed to analyze whether load mutation occurs, and quickly respond to load mutation and process load fluctuation, different learning rates are used in different load fluctuation stages, so that the model can automatically learn the characteristics of load fluctuation;

[0025] (4) realize the accurate prediction of comprehensive energy system multi-element load short-term load demand. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 It is the prediction flowchart of the present application;

[0027] Figure 2 It is the structure diagram of comprehensive energy system in the present application;

[0028] Figure 3 It is the multi-task learning model structure diagram in the present application. DETAILED DESCRIPTION

[0029] The application will be further described in detail below in connection with the embodiments, but the scope of protection of the application is not limited thereto.

[0030] The application relates to a comprehensive energy system multi-element load short-term prediction method based on multi-task learning.

[0031] In the application, the comprehensive energy system multi-element load short-term prediction method based on multi-task learning of progressive hierarchical extraction is used to analyze the coupling between different loads and balance the prediction performance of sub-tasks, and on this basis, steep load analysis is introduced to meet the demand of rapid response to load fluctuation.

[0032] The application in practical application includes the following steps:

[0033] (1) preprocessing sample data;

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

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

[0036] (4) performing load prediction by using the trained prediction model.

[0037] In the actual application process, the above steps (1) and (2) have no priority.

[0038] The steps are specifically described below in connection with the embodiments.

[0039] (1) preprocessing sample data;

[0040] In the application, the sample data is an input-output data pair selected according to a given rule, and the given rule includes:

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

[0042] b) the selected sample should contain data under various working conditions of the comprehensive energy system, cover the known fluctuation range as much as possible, and should eliminate extreme abnormal data.

[0043] The preprocessing of the sample data includes the following steps:

[0044] (1-1) collecting load data and influencing factor data of comprehensive energy system multi-element load;

[0045] The load data herein includes electrical load data, cooling load data and heating load data; of course, in actual 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] The influencing factor data refers to data that may affect the load change data, 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) Perform abnormal data processing;

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

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

[0050] The historical operation data of the 12 input variables screened and processed are collected. According to these data, the data space is divided by gridding, the mutual information value of each grid data is calculated and normalized to obtain the maximum information coefficient (MIC),

[0051]

[0052]

[0053] Wherein, 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, generally taking the 0.6 power of the total amount of data;

[0054] It is considered that when the MIC value is greater than 0.4 in the actual system operation process, there is a relatively significant correlation between the two variables, so the influencing factor data whose MIC value is greater than 0.4 with the three kinds of load is retained.

[0055] The data processed by dimensionality reduction 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] obtaining coupling correlation between the load data after dimension reduction processing, i.e., the two-two coupling between the electric load data, the cold load data and the thermal load data, by using the prediction model;

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

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

[0059] The prediction model comprises an input layer, a coupling analysis-multigranularity feature extraction layer, a gate layer and an output layer arranged in sequence; the coupling analysis-multigranularity feature extraction layer is arranged based on a bidirectional long short-term memory network.

[0060] The coupling analysis-multigranularity feature extraction layer comprises a shared layer and one or more expert layers in parallel, any expert layer matches the type of sample data; the gate layer matched with any expert layer is provided with a corresponding gate unit, and the shared layer is arranged in cooperation with all gate units.

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

[0062] In the coupling analysis-multigranularity feature extraction layer, the shared layer is used to mine the coupling between loads and capture the overall time sequence features of the input data, which is composed of a bidirectional long short-term memory network, and the number of neurons is set to 128.

[0063] In the coupling analysis-multigranularity feature extraction layer, the expert layer is used to extract more fine and specific features for a specific task, and is used to optimize the demand for a specific task, and is also composed of a bidirectional long short-term memory network, taking the three loads predicted as three sub-tasks as an example, here 3 expert layers are provided, which correspond to the types of sample data, i.e., electric load data, cold load data and thermal load data; wherein, because the cold and thermal loads have similar trends in load demand and fluctuation, the number of experts of the expert layers corresponding to the electric load, the cold load and the thermal load is set to 6, 3 and 3 respectively, and the number of neurons of each expert layer is set to 32; by setting separate expert layers, the performance of the sub-tasks can be balanced while ensuring the prediction accuracy; the data output by the input layer is input to the expert layers and the shared layer at the same time;

[0064] The coupling analysis multi-granularity feature extraction layer is a progressive hierarchical extraction manner, and the multi-task learning model represented not only has a parameter sharing layer, but also has a corresponding expert task layer for each sub-task, which enables the multi-energy system multi-element load short-term prediction to not only share parameters and learn the coupling between load data through the sharing layer, but also learn different data features of different loads through the respective unique expert task layer, thereby enabling more detailed prediction of different loads;

[0065] The gating layer is used to determine how to weight and adjust the data input into the gating layer, and is constructed by a plurality of fully connected layers, wherein the output of the gating layer is as follows,

[0066]

[0067] wherein the sigma (·) activation function is a sigmoid function, W is a weight matrix of the gating layer, b is a bias of the gating layer, x load (t) is an output of the sharing layer, x i (t) is an output of the expert layer.

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

[0069] (3) inputting 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 sub-task loss function, the cold load prediction sub-task loss function and the heat load prediction sub-task loss function.

[0071] The loss functions of the three sub-tasks are as follows:

[0072]

[0073]

[0074] wherein, respectively represent the predicted values of the electric load, the cold load and the heat load, x1, x2 and x3 respectively represent the true values of the electric load, the cold load and the heat load, i is the i-th data of each load, and N is the total amount of samples;

[0075] The total task loss function satisfies,

[0076] l = alpha1L1 + alpha2L2 + alpha3L3

[0077] wherein alpha1, alpha2 and alpha3 are set to 0.4, 0.2 and 0.4 respectively according to human experience;

[0078] Find the model parameters that minimize the total task loss function through model training.

[0079] In the training process, the learning rate of the Adam optimizer is initially set to 0.001, and other parameters are default values. Further, a steep load change analysis is introduced, and different learning rates are applied to respond to load fluctuations in different stages of load change. Specifically, a load real-time sensitivity index LSSI and a standard deviation in a rolling window segment are introduced. The calculation of LSSI is based on the change of mean square error, and the length of the rolling window is set to 16 according to the fluctuation time range of the load data. The threshold values of LSSI and standard deviation are set. During the model training process, when the value of LSSI is greater than the threshold value and the standard deviation in the rolling window segment is also greater than the threshold value, it is determined that the load has a significant fluctuation, so the model learning rate is adjusted to 0.0001. When the condition is not met, the model learning rate is adjusted to 0.001, that is, the learning rate of the model under normal circumstances is 0.001.

[0080] The load mutation sensitivity index and the standard deviation in the rolling window segment are used to determine the load change state of the comprehensive energy system, and the corresponding learning rate is used to train the prediction model according to the load change state.

[0081] The load mutation sensitivity index satisfies,

[0082]

[0083] Where t1 is the time point of load mutation, t2 is 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 occurrence of load mutation, E before The mean square error after the occurrence of load mutation; the greater the load mutation sensitivity index, the faster the model can respond to load fluctuations.

[0084] By changing the learning rate of the model, the model behaves differently in different load change stages, which can make the model stable in the case of severe load change and accelerate convergence in the case of stable load, effectively improving the accuracy and stability of load prediction.

[0085] Further, the gradient descent algorithm is used to train the prediction model, update the model parameters, calculate the gradient of the loss function with respect to the parameters, and update the weight matrix and bias term W, b of the prediction model, as follows,

[0086]

[0087]

[0088] Wherein, η is the learning rate of gradient descent algorithm, W(k) is the weight matrix corresponding to the sample data at the k time, W(k-1) is the weight matrix corresponding to the sample data at the k-1 time, b(k) is the bias corresponding to the sample data at the k time, b(k-1) is the bias corresponding to the sample data at the k-1 time, The partial differential symbol is represented, and L is the total task loss function.

[0089] (4) Load prediction is performed by using the trained prediction model.

[0090] In order to realize the above-mentioned embodiment, a computer readable storage medium is provided, which stores a multi-task learning based comprehensive energy system multi-element load short-term prediction program on it. When the multi-task learning based comprehensive energy system multi-element load short-term prediction program is executed by a processor, the multi-task learning based comprehensive energy system multi-element load short-term prediction method is realized.

[0091] In order to realize the above-mentioned embodiment, a computer device is also provided, which includes a memory, a processor and a computer program stored on the memory and executable on the processor. When the processor executes the program, the multi-task learning based comprehensive energy system multi-element load short-term prediction method is realized.

[0092] The multi-task learning based comprehensive energy system multi-element load short-term prediction method in practical application includes the following processes:

[0093] Parameter initialization: set appropriate initial values for the hyperparameters in the method, import the dimensionality-reduced comprehensive energy system multi-element load data and sample data of influencing factors in the import interface of the control computer model, and divide the training set, validation set and test set of the model; and set the prediction time length of H hours in the future;

[0094] Offline training: first, input the algorithm in the form of code into the control computer to preliminarily realize the building of the model framework, then input the dimensionality-reduced comprehensive energy system multi-element load data and sample data of influencing factors into the algorithm to train the algorithm to obtain the final appropriate hyperparameter value, and finally obtain the trained algorithm model;

[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 length H. By measuring the variable data of the comprehensive energy system online, and executing the corresponding program of the algorithm, the various energies of the comprehensive energy system can be predicted in real time.

[0096] The multi-element load short-term prediction method, medium, device and application of the integrated energy system overcome the problems of low prediction accuracy, difficulty in mining load coupling and high economic cost of the existing integrated energy system multi-element load short-term prediction, are easy to operate and have interpretability.

[0097] The above merely describes preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacements or changes according to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A short-term forecasting method for multiple loads in an integrated energy system based on multi-task learning, characterized in that: A prediction model based on a bidirectional long short-term memory network (LSTM) for multi-task learning is constructed. The prediction model includes an input layer, a coupling analysis-multi-granularity feature extraction layer, a gating layer, and an output layer arranged sequentially. The coupling analysis-multi-granularity feature extraction layer is based on the bidirectional LTM network and includes a shared layer and one or more expert layers, with each expert layer matching the type of the sample data. The gating layer, which is configured to work with any of the expert layers, contains a corresponding gating unit. The shared layer is configured to work with all the gating units. The sample data includes load data and influencing factor data of the multi-load of the integrated energy system. The prediction model is trained with the preprocessed sample data, and the weights and biases of each network layer in the prediction model are updated. The trained prediction model is then used to make short-term predictions of the multi-load of the integrated energy system.

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

3. The method for short-term forecasting of multiple loads in an integrated energy system based on multi-task learning according to claim 1, characterized in that: The number of types of load data corresponds to the number of expert layers.

4. The method for short-term forecasting of multiple loads in an integrated energy system based on multi-task learning according to claim 3, characterized in that: The load data includes electrical load data, cooling load data, and heating load data. The coupling relationship between the load data is the pairwise coupling between electrical load data, cooling load data, and heating load data.

5. The method for short-term forecasting of multiple loads in an integrated energy system based on multi-task learning according to claim 4, characterized in that: The multi-task loss function of the prediction model is positively correlated with the loss functions of the electrical load prediction sub-task, the cooling load prediction sub-task, and the heating load prediction sub-task.

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

7. The method for short-term forecasting of multiple loads in an integrated energy system based on multi-task learning according to claim 6, characterized in that: The load mutation sensitivity index is satisfied. ， in, The time point at which the load mutation occurs. To determine the time point where the error is less than the preset value, Mean square error before the load mutation occurs Mean square error after a sudden change in load.

8. 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 dimensionality reduction includes local temperature data, local air pressure data, local humidity data, and local dew point temperature data.

Images