Comprehensive energy multi-element load short-term prediction method based on secondary decoupling

Abstract

The invention provides a comprehensive energy multi-element load short-term prediction method based on secondary decoupling, and the method comprises the steps: collecting a multi-element load data set, reconstructing a winter data set and a summer data set in the multi-element load data set, and carrying out the primary decoupling; decomposing the reconstructed winter data set and summer data set; dividing the decomposed data into a periodic term and a trend term so as to perform secondary decoupling; screening the trend items, and predicting the screened trend items to obtain a trend item prediction result; and performing multivariate load prediction on the periodic term to obtain a periodic term prediction result, and adding the trend term prediction result and the periodic term prediction result to obtain a final prediction result. According to the method, the coupling of the multi-element load piece can be effectively captured, accurate decomposition of nonlinear and non-stationary signals can be realized, the anti-noise capability and the time frequency resolution are improved, and the balance is considered while the prediction accuracy is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of comprehensive energy forecasting, and in particular to a comprehensive energy multi-load short-term forecasting method based on secondary decoupling. Background Art

[0002] In the traditional prediction of multi-load, it is easy to produce biased prediction results, such as greatly improving the prediction accuracy of a certain load, but sacrificing the accuracy of other load predictions, resulting in an imbalance in the final prediction results. In order to solve the problem of prediction imbalance in multi-load forecasting, researchers proposed a multi-task learning (MTL) model. MTL trains multiple related tasks simultaneously by sharing the underlying network structure, in order to improve the performance of the model on all tasks. However, MTL also has a limitation, that is, it incorporates all input influencing factors into the model, but cannot distinguish whether these factors are positive or negative for the multi-load forecasting results, so it does not fundamentally solve the problem of prediction imbalance. With the deepening of research, the Mixture of experts (MMoE) model was proposed. MMoE assigns a dedicated expert network to each task and uses a gating mechanism to flexibly control the output combination of different experts to achieve the best adaptation to each task. This innovation effectively alleviates the problem of negative transfer between tasks, that is, the performance improvement of one task is no longer at the expense of the performance of another task, which makes MMoE show significant advantages in multi-load forecasting. However, despite the many improvements brought by the MMoE model, its network construction and parameter interaction mechanism are new challenges. Researchers have begun to explore improving the model. For example, by optimizing the expert selection strategy, improving the gating mechanism, or designing a more efficient network structure, researchers are committed to breaking through the limitations of MMoE in terms of network construction and parameter interaction mechanism. These improvements not only help improve model performance, but also reduce the consumption of computing resources, making multi-task learning more efficient and feasible in practical applications.

[0003] There are two common challenges in the field of multi-task learning. The phenomenon of unbalanced performance between tasks. In multi-task learning, when the correlation between tasks is complex, it may happen that improving the effect of one task will be accompanied by a decrease in the effect of other tasks. In the IES multi-load forecasting task, the improvement of the forecasting effect of one load is often accompanied by a decrease in the forecasting effect of other loads. For example, the prediction accuracy of electric load is improved, while the prediction accuracy of cold load and heat load is reduced. The phenomenon of negative transfer. Negative transfer refers to the decline in learning performance when the correlation between tasks is not strong or even conflicts in multi-task learning. The coupling characteristics between the loads of IES are constantly changing. For example, the coupling degree between the loads is strong in winter and autumn. In summer, the heat load fluctuates greatly and its correlation with the cold load and electric load is weak, which leads to the phenomenon of negative transfer. The difference between the two is: First, the phenomenon of unbalanced performance between tasks means that while improving the performance of one task, the performance of other tasks may be sacrificed. It is impossible to achieve the ideal effect of each task at the same time. Secondly, the phenomenon of negative transfer means that the effect of multi-task learning is not as good as the effect of training each task separately, and it is not as good as predicting each load independently. Training the model together will lead to a decrease in the effect of the task. That is to say, when the correlation between loads or between two loads is not strong, information sharing between loads through multi-task learning will affect the performance of the network and lead to a decrease in the prediction accuracy of each load. This phenomenon is more obvious when the correlation between tasks is not strong or even when there is a conflict. MMOE uses a gated network combined with an expert network to deal with the correlation of tasks. Although it alleviates the negative transfer phenomenon to a certain extent, it ignores the differences and interactive information between expert networks, and its potential needs to be further explored. In order to solve the above problems, the existing technology proposes a method of using prior knowledge to capture the correlation of complex tasks, using shared experts and task-specific experts to reduce the interference of harmful parameters in the information, and using multiple layers of experts and gated networks in the form of Progressive Layered Extraction (PLE) to extract deeper information from the underlying expert network and separate task parameters from the high-level expert network. Since the cooling load, heating load and electric load of the integrated energy system contain a large number of coupling characteristics and seasonal characteristics, the time series of IES multivariate loads can be segmented into many data blocks through PatchTST data segmentation, and each data block can be reconstructed through mask capture. Since the reconstructed data accepts the seasonal and coupled correlation structure, the trend items and seasonality are more prominent. Existing studies have shown that the decomposition and aggregation of time series can help extract deep trends and seasonal changes in sequence data. A new and more effective PatchTST-PLE hybrid prediction model is provided for the integrated energy system.

[0004] In summary, it is difficult to effectively capture the coupling between multiple loads in the existing technology, and the noise resistance and time frequency resolution are poor, and the prediction accuracy is poor, and the accuracy and balance of the comprehensive energy short-term multiple load forecasting technology are insufficient. Summary of the invention

[0005] Based on this, the purpose of the present invention is to provide a comprehensive energy multi-load short-term forecasting method based on secondary decoupling to solve the deficiencies in the above-mentioned prior art.

[0006] The present invention provides a method for short-term forecasting of comprehensive energy multi-load based on secondary decoupling, the method comprising: Collecting a multivariate load data set, and dividing the multivariate load data set into a first half data set and a second half data set according to seasonal characteristics and data coupling relationships, and reconstructing a winter data set and a summer data set in the multivariate load data set based on the first half data set and the second half data set to perform a decoupling; Decomposing the reconstructed winter data set and the reconstructed summer data set based on TTAO-VMD to obtain decomposed data; Dividing the decomposed data into periodic items and trend items by fine composite multi-scale sample entropy to perform secondary decoupling; The trend items are screened by using MIV feature screening, and the trend items after screening are predicted by using Lasso linear regression model to obtain trend item prediction results; Progressive Layered Extraction multi-task learning is used to perform multivariate load forecasting on the periodic item to obtain a periodic item prediction result, and the trend item prediction result and the periodic item prediction result are added together to obtain a final prediction result.

[0007] Compared with the prior art, the beneficial effects of the present invention are as follows: by reconstructing the winter data set and the summer data set, the first coupling is performed, and by dividing the decomposed data into periodic terms and trend terms for secondary decoupling, the coupling of multivariate load components can be effectively captured, and the reconstructed data is decomposed by TTAO-VMD, which can achieve accurate decomposition of nonlinear and non-stationary signals, improve noise resistance and time-frequency resolution, and divide by fine composite multi-scale sample entropy, which can reduce feature redundancy and data noise in the feature extraction process and improve prediction accuracy, and multivariate load prediction of periodic terms through Progressive Layered Extraction multi-task learning can effectively improve prediction accuracy while taking into account balance.

[0008] Furthermore, the step of dividing the multivariate load data set into a first half data set and a second half data set according to seasonal characteristics and data coupling relationships includes: Dividing the load sequence in the multivariate load data set into a plurality of subsequences; A random mask is used to cover part of the data in some of the subsequences to capture the coupling relationship between the multivariate loads for feature extraction and reconstruction.

[0009] Furthermore, the step of reconstructing the winter data set and the summer data set in the multivariate load data set based on the first half data set and the second half data set includes: The multivariate load data set includes spring data, summer data, autumn data and winter data, the spring data, summer data, autumn data and winter data are arranged, and the spring data, summer data, autumn data and winter data are divided into first half data and second half data sets according to seasonal characteristics and data coupling relationships; Extracting local information in the multivariate load sequence in the multivariate load data set based on PtachTST, and mining the time dependency of the multivariate load sequence through a multi-head attention mechanism to obtain a feature vector; Obtaining attention distributions of different subspaces of the multivariate load sequence based on a plurality of independent attention mechanisms; Performing bidirectional modeling on the feature vector based on BiLSTM; Training PatchTST-BiLSTM based on the training set to obtain a trained PatchTST-BiLSTM, and adjusting hyperparameters on the training set to obtain a load reconstruction result under an optimal hyperparameter configuration; The trained PatchTST-BiLSTM is reconstructed based on the test set to obtain the final IES multi-cloud load reconstruction result.

[0010] Furthermore, the step of decomposing the reconstructed winter data set and the reconstructed summer data set based on TTAO-VMD includes: Introducing dynamic tuning mechanism into TTAO algorithm; The reconstructed winter data set and the reconstructed summer data set are decomposed according to the TTAO algorithm after the dynamic tuning mechanism is introduced.

[0011] Furthermore, the adjustment formula of the center frequency of the TTAO algorithm is: ; In the formula, represents the center frequency, Indicates frequency components, Indicates that at time step hour, Indicates that at time step The The center frequency of a frequency component, represents the partial derivative, represents the objective function in the decomposition process, represents the step size of the center frequency; The adjustment formula of the bandwidth parameter of the TTAO algorithm is: ; In the formula, Indicates bandwidth parameter, Indicates that at time step The The value of the bandwidth parameter, Indicates the step size of the bandwidth parameter.

[0012] Furthermore, the expression of the objective function in the decomposition process is: ; In the formula, The expression representing the objective function in the decomposition process, Indicates The intrinsic mode functions, for The derivative of represents the frequency bandwidth of the signal, Indicates that at time step The The value of the bandwidth parameter. Furthermore, the step of dividing the decomposed data into periodic items and trend items by fine composite multi-scale sample entropy includes: Selecting a similarity threshold for the decomposed data based on fine composite multi-scale sample entropy and using a fine-grained scale, and evaluating it in combination with several entropy metrics; Fine scale decomposition is introduced to capture the complex dynamic features in the decomposed data, so as to divide the decomposed data into periodic items and trend items.

[0013] Furthermore, the step of using MIV feature screening to screen the trend items and using Lasso linear regression model to predict the screened trend items includes: The meteorological data and the numerical values ​​of the calendar rules in the trend item are disturbed by the MIV model based on the electric load, the heating load and the cooling load to obtain the MIV value; The Lasso linear regression model is adopted and the preset features in the MIV value are eliminated by minimizing the loss function of the L1 norm penalty.

[0014] Furthermore, the expression of the loss function that minimizes the L1 norm penalty is: ; In the formula, represents the loss function that minimizes the L1 norm penalty, Indicates The actual value of the samples, Indicates The predicted value of samples, represents the regression coefficient of the MIV model, Indicates The regression coefficient corresponding to the feature is represents the number of samples, represents the number of features, represents the regularization parameter.

[0015] Furthermore, the step of using Progressive Layered Extraction multi-task learning to perform multivariate load forecasting on the periodic item to obtain a periodic item prediction result includes: Based on Progressive Layered Extraction, a multi-layer expert network and a gating network are used to extract the depth information in the periodic term; A fully connected layer is used as a learnable weight matrix, and a Softmax function is used to assign weights to each layer of the expert network; The depth information extracted by each layer of the expert network is integrated to output a periodic prediction result. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of a method for short-term forecasting of integrated energy multi-loads based on secondary decoupling in an embodiment of the present invention; Figure 2 It is a structural schematic diagram of a graph network model in an embodiment of the present invention; Figure 3 It is a schematic diagram of using Lasso linear regression model prediction and Progressive Layered Extraction multi-task learning prediction in an embodiment of the present invention; Figure 4 Schematic diagram of the Progressive Layered Extraction prediction model in an embodiment of the present invention.

[0017] The following specific implementation manner will further illustrate the present invention in conjunction with the above-mentioned drawings. DETAILED DESCRIPTION

[0018] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. Several embodiments of the present invention are given in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.

[0019] It should be noted that when an element is referred to as being "fixed to" another element, it may be directly on the other element or there may be a central element. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be a central element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which the present invention belongs. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.

[0021] See also Figure 1 as well as Figure 2 , which shows a short-term forecasting method for comprehensive energy multi-load based on secondary decoupling in the first embodiment of the present invention, and the method includes steps S1 to S5: S1, collecting a multivariate load data set, and dividing the multivariate load data set into a first half data set and a second half data set according to seasonal characteristics and data coupling relationships, and reconstructing a winter data set and a summer data set in the multivariate load data set based on the first half data set and the second half data set to perform a decoupling; Specifically, the step S1 includes steps S11 to S18: S11, dividing the load sequence in the multivariate load data set into a plurality of subsequences; S12, using a random mask to mask part of the data in the plurality of subsequences to capture the coupling relationship between the multivariate loads for feature extraction and reconstruction; It can be understood that the load sequence in the multivariate load data set is divided into several subsequences (Patch), that is, a time block is regarded as a token. When the load data is segmented into Patch, two consecutive Patch areas can be overlapping or non-overlapping, which is determined by the Patch length P and the step length S (the non-overlapping area between two consecutive Patches). subsequence as token, It represents the length of the time series, and uses a random mask to cover part of the data in the patch to capture the coupling relationship between multiple loads for feature extraction and reconstruction; S13, the multivariate load data set includes spring data, summer data, autumn data and winter data, the spring data, summer data, autumn data and winter data are arranged, and the spring data, summer data, autumn data and winter data are divided into first half data and second half data sets according to seasonal characteristics and data coupling relationships; S14, extracting local information in the multivariate load sequence in the multivariate load data set based on PtachTST, and mining the time dependency of the multivariate load sequence through a multi-head attention mechanism to obtain a feature vector; S15, obtaining attention distribution of different subspaces of the multivariate load sequence based on a plurality of multi-head attention mechanisms; S16, performing bidirectional modeling on the feature vector based on BiLSTM; S17, training PatchTST-BiLSTM based on the training set to obtain a trained PatchTST-BiLSTM, and adjusting hyperparameters on the training set to obtain a load reconstruction result under an optimal hyperparameter configuration; S18, reconstructing the trained PatchTST-BiLSTM based on the test set to obtain a final IES multi-cloud load reconstruction result; It can be understood that PtachTST, as an encoder, extracts local information within the multi-load sequence, and the multi-head attention mechanism further explores the temporal dependencies of the multi-loads. By running multiple independent attention mechanisms in parallel to obtain the attention distribution of different subspaces of the input sequence, the model's ability to understand and capture complex dependencies is improved; It should be noted that BiLSTM, as a decoder, performs bidirectional modeling on the feature vector obtained by the encoder, trains PatchTST-BiLSTM based on the training set, and adjusts the hyperparameters on the training set to obtain the load reconstruction result under the optimal hyperparameter configuration. The IES multivariate load forecasting model PatchTST-BiLSTM under the optimal hyperparameter configuration is reconstructed on the test set to obtain the final IES multivariate load reconstruction result.

[0022] S2, decomposing the reconstructed winter data set and the reconstructed summer data set based on TTAO-VMD to obtain decomposed data; Specifically, the step S2 includes steps S21 to S22: S21, introduce dynamic tuning mechanism into TTAO algorithm; S22, decomposing the reconstructed winter data set and the reconstructed summer data set according to the TTAO algorithm after the dynamic tuning mechanism is introduced; It should be explained that the TTAO algorithm is based on similar triangles. With iterations, new vertices are continuously generated in the search space and used to form similar triangles of different sizes. Each triangle is regarded as a basic evolutionary unit with four bodies, namely the three vertices of the triangle and an internal random vertex. By dynamically adjusting the parameters, the decomposition process can better adapt to the time-varying characteristics and complexity of the signal. The TTAO algorithm introduces a dynamic tuning mechanism so that the tuning parameters can be adaptively adjusted as the signal changes, thereby improving the accuracy and stability of the VMD decomposition results. It should be noted that, in this embodiment, the adjustment formula of the center frequency of the TTAO algorithm is: ; In the formula, represents the center frequency, Indicates frequency components, Indicates that at time step hour, Indicates that at time step The The center frequency of a frequency component, represents the partial derivative, represents the objective function in the decomposition process, represents the step size of the center frequency; The adjustment formula of the bandwidth parameter of the TTAO algorithm is: ; In the formula, Indicates bandwidth parameter, Indicates that at time step The The value of the bandwidth parameter, Indicates the step size of the bandwidth parameter.

[0023] The expression of the objective function in the decomposition process is: ; In the formula, The expression representing the objective function in the decomposition process, Indicates The intrinsic mode functions, for The derivative of represents the frequency bandwidth of the signal, Indicates that at time step The The value of the bandwidth parameter. It is worth noting that during the optimization process, the TTAO optimization algorithm dynamically adjusts the step size according to the current signal characteristics. and : ; ; In the formula, , They respectively represent the parameters for controlling the adjustment range of the center frequency and the parameters for controlling the adjustment range of the bandwidth parameter.

[0024] S3, dividing the decomposed data into a periodic term and a trend term by fine composite multi-scale sample entropy to perform secondary decoupling; Specifically, the step S3 includes steps S31 to S32: S31, selecting a similarity threshold for the decomposed data based on fine composite multi-scale sample entropy and using a fine-grained scale, and evaluating it in combination with several entropy metrics; S32, introducing fine scale decomposition to capture the complex dynamic features in the decomposed data, so as to divide the decomposed data into periodic items and trend items; It can be understood that the fine composite multi-scale sample entropy can more accurately evaluate the complexity of time series data by using finer-grained scale decomposition, optimizing the selection of similarity threshold r, and combining multiple entropy metrics to enhance the stability of the results. It introduces fine scale decomposition to obtain more accurate entropy values, thereby better capturing the complex dynamic characteristics in the data. By optimizing the scale decomposition and sample entropy calculation process, the load data is divided into periodic and trend components.

[0025] It is worth noting that the over-fine composite multi-scale sample entropy divides the decomposed data into periodic terms and trend terms. The data shows a relatively obvious regularity, the trend term component is relatively stable, and the periodic regularity and change pattern of each load in the periodic term are complex, so it is the second decoupling.

[0026] S4, using MIV feature screening to screen the trend items, and using Lasso linear regression model to predict the trend items after screening to obtain trend item prediction results; Specifically, the step S4 includes step S41 to step S42: S41, using the MIV model and based on the electric load, the heat load and the cooling load, the meteorological data in the trend item and the numerical value of the calendar rule are disturbed to obtain the MIV value; S42, using a Lasso linear regression model and eliminating preset features in the MIV value by minimizing a loss function of an L1 norm penalty; It is understandable that there are too many influencing factors in the trend item, including weather factors and calendar rules, which may have a counterproductive effect on the final prediction results. The MIV model is added for feature screening. The MIV model is based on electric load, heat load, and cooling load. By perturbing the numerical values ​​of the input meteorological data and calendar rules, the MIV value of each input feature is obtained. The absolute value of the MIV feature indicates the importance of the input feature to the output variables of electric load, cooling load, and heat load. The Lasso regression model effectively compresses the model coefficients by minimizing the loss function containing the L1 norm penalty, and automatically eliminates unimportant features, thereby constructing a more concise and easy-to-interpret model and reducing the risk of overfitting. In this embodiment, the expression of the loss function for minimizing the L1 norm penalty is: ; In the formula, represents the loss function that minimizes the L1 norm penalty, Indicates The actual value of the samples, Indicates The predicted value of samples, represents the regression coefficient of the MIV model, Indicates The regression coefficient corresponding to the feature is represents the number of samples, represents the number of features, Represents the regularization parameter, which is used to control the strength of the L1 regularization term.

[0027] It is worth noting that after the secondary decoupling, the relatively stable trend item reflects the general trend of load changes, has strong regularity, and can obtain more accurate prediction results without considering the coupling between loads. MIV feature screening is used to further enhance the accuracy of the prediction. Therefore, the Lasso linear regression model is used for prediction.

[0028] S5, using Progressive Layered Extraction multi-task learning to perform multivariate load forecasting on the periodic item to obtain a periodic item forecast result, and adding the trend item forecast result and the periodic item forecast result to obtain a final forecast result; It is worth noting that for the periodic component with strong non-stationary characteristics, it represents the periodic change law of each load and the change pattern is relatively complex. It is necessary to comprehensively consider the complex coupling relationship between loads and the independent characteristics of each load itself, and use Progressive Layered Extraction multi-task learning for multi-load forecasting.

[0029] Specifically, the step S5 includes steps S51 to S53: S51, extracting depth information in the periodic item based on Progressive Layered Extraction and using a multi-layer expert network and a gating network; S52, using a fully connected layer as a learnable weight matrix, and assigning weights to each layer of the expert network through a Softmax function; S53, integrating the depth information extracted by each layer of the expert network of the weighted candidate to output a periodic prediction result; It is understandable that shared experts and task-specific experts are used to reduce the interference of harmful parameters in common and specific information. Progressive Layered Extraction uses multi-layer experts and gated networks to extract more in-depth information from the underlying expert network and better separate task-specific parameters at the high level. The fully connected layer is used as a learnable weight matrix, and then the weight is assigned to each expert subnetwork through the Softmax function. The output tower of each task integrates the information to output the prediction results of each task. Finally, the trend item prediction results and the period item prediction results are added to obtain the final prediction results, as shown in the following figure. Figure 3 and Figure 4 shown.

[0030] In the specific implementation, the IES electricity, cooling and heating load data of a certain area from 00:00 on January 1, 2020 to 23:00 on December 31, 2020 were collected, with a data granularity of 1h and a data length of 8784. The data was divided into the first and second half of the year, and put into PatchTST respectively. According to the coupling and seasonal characteristics between the data, 2000 data were reconstructed respectively. Then, the reconstructed data and the original data of several days were used to form the complete January, February and March data as the winter training set, and the July, August and September data as the summer training set to train the model. The data distribution of February and August 2021 was used as the winter test set and the summer test set to learn the comprehensive energy consumption pattern in different seasons. The performance of the model was evaluated through the test set, and the results of the test set were used as the prediction accuracy. Meteorological data were collected from an official website, including temperature, wind speed, relative humidity, direct normal radiation (DNI), global horizontal radiation (GHI), dew point temperature and air pressure, and five calendar rules, namely month, day, hour, week and whether it is a holiday, were considered as input features. In order to comprehensively evaluate the prediction results of the model, the root mean square error (RMSE) and the coefficient of determination ( ), and the mean absolute percentage error (MAPE) is used as the evaluation index for each load.

[0031]

[0032]

[0033]

[0034] In the formula, , , They represent the root mean square error, determination coefficient, and mean absolute error percentage, respectively. represents the sample size, that is, the total number of observations or data points in the data set. Represents the predicted value of the T-th sample, that is, the prediction result of the model for the T-th data point. represents the actual value of the T-th sample, that is, the true observation result of the T-th data point, T Represents the index of the sample, that is, the Tth data point in the data set. At the same time, a balanced accuracy (BMA) is defined to measure the multivariate load as a whole, and the model is evaluated by taking into account both accuracy and balance. BMA is defined as: ; ; In the formula, , , They are the MAPE of electric load, MAPE of cooling load, and MAPE of heating load. , , They are the weight of electric load, cooling load and heat load respectively. Taking into account the accuracy and balance of the research, the weight coefficients in the formula are 0.3, 0.4 and 0.3 respectively for winter prediction and 0.3, 0.3 and 0.4 respectively for summer prediction. The smaller the value, the better the model performance.

[0035] In order to fully reflect the coupling between data and give full play to the performance of the program, at the beginning of the experiment, the three loads of electricity, cold and heat are all converted into KW. The conversion formula is: ; ; After the experiment, convert it to the original unit for performance evaluation.

[0036] The specific implementation process is as follows: first, the original multivariate load sequence is processed in blocks, and the original IES multivariate load sequence is divided into several subsequences (Patch). Then, each Patch is regarded as a token, and three encoding operations are performed. They are tag encoding, position encoding, and time encoding. Among them, tag encoding performs preliminary time feature extraction on each token based on the time convolution layer, position encoding uses the sine and cosine encoding method of Transformer to give position information to each token, and time encoding is to provide the model with the timestamp information of the data. The encoded independent tokens are sent to the Transformer encoder, and a global dependency relationship is established based on the Transformer's multi-head attention mechanism. The local information in each token is extracted nonlinearly through the position feedforward fully connected neural network. Finally, BiLSTM is used to capture the bidirectional time dependency relationship, which acts as a decoder, and the reconstruction result is output by projecting it to the prediction window dimension through the fully connected layer.

[0037] When performing Patch segmentation on the payload data, two consecutive Patch areas can be overlapping or non-overlapping, depending on the length of the Patch. and step length (the non-overlapping area between two consecutive patches). Finally, the load sequence is divided into subsequences as tokens for feature extraction. This method of dividing the load data into subsequences, on the one hand, reduces the number of input tokens by the sequence length. Reduce to , reducing the time and space complexity of the traditional Transformer. On the other hand, the model can model longer historical load sequences and retain local information of the load. Multiple experiments have shown that when the load input sequence length is the past six months, the length of the patch The step length is 10. The effect is better when it is 10. That is, the blocks are processed in a non-overlapping manner, each Patch is a 10-hour local load sequence information, the three columns of electric load, cooling load, and heating load are coupled and reconstructed as input series, and the remaining columns are feature series. Each Patch constitutes the global multivariate load sequence information for the past six months.

[0038] Traditional signal decomposition methods such as VMD use fixed tuning parameters for decomposition, which may lead to a decrease in decomposition accuracy when processing non-stationary signals. The TTAO optimization algorithm introduces a dynamic tuning mechanism so that the tuning parameters can be adaptively adjusted as the signal changes, thereby improving the accuracy and stability of the decomposition results. The core of the TTAO optimization algorithm is to dynamically adjust the parameters in the decomposition process through time-varying parameter tuning and adaptive optimization strategies.

[0039] Assume that the decomposition , the TTAO optimization algorithm dynamically adjusts the center frequency and bandwidth parameters To optimize the decomposition process. The adjustment formula of the center frequency is: ; The adjustment formula of bandwidth parameters is: ; in, is the objective function in the decomposition process (such as the reconstruction error of the signal or the frequency bandwidth constraint), and They are the step sizes of the center frequency and bandwidth parameters, respectively, which are adaptively adjusted according to the time-varying characteristics of the signal.

[0040] The objective function of the TTAO optimization algorithm can be to minimize the signal reconstruction error ,in, It is Intrinsic Mode Functions (IMFs), yes The derivative of represents the frequency bandwidth of the signal.

[0041] During the optimization process, the TTAO optimization algorithm dynamically adjusts the step size according to the current signal characteristics. and : ; ; in, and is the parameter that controls the step size adjustment amplitude. The function represents the sign of the derivative, which is used to control the increase or decrease of the step size. Then, the refined composite multiscale sample entropy (Refined Composite Multiscale SampleEntropy) is introduced. The refined composite multiscale sample entropy optimizes the similarity threshold by using a finer-grained scale decomposition. The choice of multiple entropy measures is combined to enhance the stability of the results and to more accurately evaluate the complexity of time series data.

[0042] Progressive Layered Extraction divides the expert subnet into two types: one is a shared expert subnet expert (Shared), and the other is a separate expert subnet expert for each subtask (Task-Specific). Therefore, this design retains the transfer learning ability between tasks by sharing the expert subnet, and can avoid the noise interference of harmful parameters when the correlation between different tasks is weak, thereby avoiding the negative transfer phenomenon.

[0043] Considering that there are too many influencing factors of weather factors and calendar rules in the trend item, which may have a counter-effect on the final prediction results, the MIV model is added for feature screening. The MIV model is based on electric load, heat load, and cooling load. By perturbing the numerical values ​​of the input meteorological data and calendar rules, the MIV value of each input feature is obtained. Its absolute value indicates the importance of the input feature to the output variables of electric load, cooling load, and heat load. The Lasso regression model effectively compresses the model coefficients by minimizing the loss function containing the L1 norm penalty, and automatically eliminates unimportant features, thereby constructing a more concise and easy-to-interpret model. This sparsity not only improves the generalization ability of the model, but also reduces the risk of overfitting.

[0044] The comparison results of the prediction models are shown in Table 1: Table 1

[0045] It can be seen from Table 1 that the proposed model has the highest accuracy in predicting heating load, cooling load and electrical load in both winter and summer, and the value of the mean absolute percentage error of the final total balance accuracy is also the smallest, which fully demonstrates that the method proposed in this paper has the best effect in terms of prediction accuracy and balance.

[0046] In summary, the comprehensive energy multi-load short-term forecasting method based on secondary decoupling in the above-mentioned embodiment of the present invention applies the PatchTST model to multi-load forecasting, and effectively captures the coupling between multi-loads by dividing the time series into blocks and improving the model's ability to handle long-term dependencies. It also reconstructs part of the data considering the seasonal characteristics of meteorological factors and calendar rules, thereby achieving significant performance improvement in multivariate time series forecasting tasks; the TTAO-VMD algorithm effectively improves the time-frequency analysis performance in signal processing by combining the triangular topology aggregation optimizer with the variational mode decomposition technology, realizes the accurate decomposition of nonlinear and non-stationary signals, and demonstrates excellent noise resistance and time-frequency resolution; a feature processing method is proposed, which divides the input features into two groups for differential processing according to the difference in fine composite multi-scale sample entropy (fine composite multi-scale sample entropy) to pay more attention to historical load data. Through this method, feature redundancy and data noise in the feature extraction process can be reduced, thereby improving the prediction accuracy of the model; an improved multi-task learning PatchTST-PLE model is constructed. In MMoE, multiple tasks share all experts, and different tasks control different expert coefficients through gate weights. In PatchTST-PLE, in addition to shared experts, multiple tasks also have private experts and consider the interactive characteristics of different experts, thereby achieving improved prediction accuracy while taking into account balance.

[0047] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0048] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A short-term forecasting method for comprehensive energy multi-load based on secondary decoupling, characterized in that: The method comprises: Collecting a multivariate load data set, and dividing the multivariate load data set into a first half data set and a second half data set according to seasonal characteristics and data coupling relationships, and reconstructing a winter data set and a summer data set in the multivariate load data set based on the first half data set and the second half data set to perform a decoupling; Decomposing the reconstructed winter data set and the reconstructed summer data set based on TTAO-VMD to obtain decomposed data; Dividing the decomposed data into periodic items and trend items by fine composite multi-scale sample entropy to perform secondary decoupling; The trend items are screened by using MIV feature screening, and the trend items after screening are predicted by using Lasso linear regression model to obtain trend item prediction results; Progressive Layered Extraction multi-task learning is used to perform multivariate load forecasting on the periodic item to obtain a periodic item prediction result, and the trend item prediction result and the periodic item prediction result are added together to obtain a final prediction result.

2. The method for short-term forecasting of comprehensive energy multi-load based on secondary decoupling according to claim 1 is characterized in that: The step of dividing the multivariate load data set into a first half data set and a second half data set according to seasonal characteristics and data coupling relationships comprises: Dividing the load sequence in the multivariate load data set into a plurality of subsequences; A random mask is used to cover part of the data in some of the subsequences to capture the coupling relationship between the multivariate loads for feature extraction and reconstruction.

3. The method for short-term forecasting of comprehensive energy multi-load based on secondary decoupling according to claim 1 is characterized in that: The step of reconstructing the winter data set and the summer data set in the multivariate load data set based on the first half data set and the second half data set comprises: The multivariate load data set includes spring data, summer data, autumn data and winter data, the spring data, summer data, autumn data and winter data are arranged, and the spring data, summer data, autumn data and winter data are divided into first half data and second half data sets according to seasonal characteristics and data coupling relationships; Extracting local information in the multivariate load sequence in the multivariate load data set based on PtachTST, and mining the time dependency of the multivariate load sequence through Multi-Head Attention to obtain a feature vector; Obtaining attention distributions of different subspaces of the multivariate load sequence based on a plurality of independent attention mechanisms; Performing bidirectional modeling on the feature vector based on BiLSTM; Training PatchTST-BiLSTM based on the training set to obtain a trained PatchTST-BiLSTM, and adjusting hyperparameters on the training set to obtain a load reconstruction result under an optimal hyperparameter configuration; The trained PatchTST-BiLSTM is reconstructed based on the test set to obtain the final IES multi-cloud load reconstruction result.

4. The method for short-term forecasting of comprehensive energy multi-load based on secondary decoupling according to claim 1 is characterized in that: The step of decomposing the reconstructed winter data set and the reconstructed summer data set based on TTAO-VMD comprises: Introducing dynamic tuning mechanism into TTAO algorithm; The reconstructed winter data set and the reconstructed summer data set are decomposed according to the TTAO algorithm after the dynamic tuning mechanism is introduced.

5. The method for short-term forecasting of comprehensive energy multi-load based on secondary decoupling according to claim 4 is characterized in that: The adjustment formula of the center frequency of the TTAO algorithm is: ； In the formula, represents the center frequency, Indicates frequency components, Indicates that at time step hour, Indicates that at time step The The center frequency of a frequency component, represents the partial derivative, represents the objective function in the decomposition process, represents the step size of the center frequency; The adjustment formula of the bandwidth parameter of the TTAO algorithm is: ； In the formula, Indicates bandwidth parameter, Indicates that at time step The The value of the bandwidth parameter, Indicates the step size of the bandwidth parameter.

6. The method for short-term forecasting of comprehensive energy multi-load based on secondary decoupling according to claim 5 is characterized in that: The expression of the objective function in the decomposition process is: ； In the formula, The expression representing the objective function in the decomposition process, Indicates The intrinsic mode functions, for The derivative of represents the frequency bandwidth of the signal, Indicates that at time step The The value of the bandwidth parameter.

7. The method for short-term forecasting of comprehensive energy multi-load based on secondary decoupling according to claim 1 is characterized in that: The step of dividing the decomposed data into periodic items and trend items by fine composite multi-scale sample entropy comprises: Selecting a similarity threshold for the decomposed data based on fine composite multi-scale sample entropy and using a fine-grained scale, and evaluating it in combination with several entropy metrics; Fine scale decomposition is introduced to capture the complex dynamic features in the decomposed data, so as to divide the decomposed data into periodic items and trend items.

8. The method for short-term forecasting of comprehensive energy multi-load based on secondary decoupling according to claim 1 is characterized in that: The step of using MIV feature screening to screen the trend items and using Lasso linear regression model to predict the screened trend items includes: The meteorological data and the numerical values ​​of the calendar rules in the trend item are disturbed by the MIV model based on the electric load, the heating load and the cooling load to obtain the MIV value; The Lasso linear regression model is adopted and the preset features in the MIV value are eliminated by minimizing the loss function of the L1 norm penalty.

9. The method for short-term forecasting of comprehensive energy multi-load based on secondary decoupling according to claim 8 is characterized in that: The expression of the loss function that minimizes the L1 norm penalty is: ； In the formula, represents the loss function that minimizes the L1 norm penalty, Indicates The actual value of the samples, Indicates The predicted value of samples, represents the regression coefficient of the MIV model, Indicates The regression coefficient corresponding to the feature is represents the number of samples, represents the number of features, represents the regularization parameter.

10. The method for short-term forecasting of comprehensive energy multi-load based on secondary decoupling according to claim 1 is characterized in that: The step of using Progressive Layered Extraction multi-task learning to perform multivariate load forecasting on the periodic item to obtain a periodic item prediction result includes: Based on Progressive Layered Extraction, a multi-layer expert network and a gating network are used to extract the depth information in the periodic term; A fully connected layer is used as a learnable weight matrix, and a Softmax function is used to assign weights to each layer of the expert network; The depth information extracted by each layer of the expert network is integrated to output a periodic prediction result.

Images