Integrated energy system load forecasting method based on multi-energy demand response and ICPO-BiLSTM model

Abstract

The load forecasting method of the integrated energy system based on multi-energy demand response and ICPO-BiLSTM model includes the following steps: based on the multi-energy demand response mechanism, extract the signal characteristics of the multi-energy demand response and construct the multi-energy demand response model; based on the chaos mapping theory and the elite reverse learning strategy, improve the crown porcupine optimization algorithm CPO; optimize the BiLSTM neural network model based on the multi-head self-attention mechanism; and jointly forecast multiple loads. This forecasting method can effectively deal with the influence of the coupling relationship between multiple loads and the user demand response on the load forecasting accuracy, and can improve the accuracy of the multiple load forecasting.

Description

Technical Field

[0001] The present invention relates to the technical field of integrated energy system load forecasting, and in particular to an integrated energy system load forecasting method based on multi-energy demand response and an ICPO-BiLSTM model. Background Art

[0002] The integrated energy system breaks the limitation of the independent operation of various energy systems in the traditional energy supply mode. On the basis of making full use of multiple energy resources, it effectively reduces the uncertainty in energy supply. Therefore, the energy supply mode gradually shifts to the integrated energy system. Load forecasting is an important prerequisite for energy management and scheduling optimization of the integrated energy system. It can guide the system to rationally utilize various energy resources and provide a basis for its subsequent efficient and stable operation. Therefore, it is of great significance to study load forecasting.

[0003] In the existing technical literature: Literature [1]: "Short-term power load forecasting based on attention mechanism optimization of long short-term memory network" (Wang Jian, Yi Shuhui, Liu Hao, et al. Short-term power load forecasting based on attention mechanism optimization of long short-term memory network [J]. Journal of South-Central University for Nationalities (Natural Science Edition), 2023, 42(01): 73-81.) proposed to use an improved AM-LSTM model to forecast the load, which improved the load forecasting accuracy. However, the above literature only considers the forecasting of power load, while the integrated energy system contains multiple types of loads, and there is a coupling relationship between these loads, which reduces the accuracy of single load forecasting.

[0004] Reference [2]: "Short-term forecasting of multi-scale loads in integrated energy systems based on multi-scale feature extraction and nonlinear fusion" (Fu Wenlong, Zhang Xuanrui, Zhang Hairong, et al. Short-term forecasting of multi-scale loads in integrated energy systems based on multi-scale feature extraction and nonlinear fusion [J]. Journal of Electric Power Systems and Automation, 2023, 35(12): 89-99.) proposes a multi-scale load forecasting model. The model extracts multi-scale features after decomposing the time series input into an embedded decomposition module, and uses an adaptive nonlinear fusion module to achieve forecast result fusion, thereby improving the generalization and forecasting performance of the model. However, the above-mentioned literature still has shortcomings in considering the relevant factors affecting the multi-scale load forecasting results, and the overall performance of the forecasting model still needs to be improved.

[0005] Reference [3]: "Short-term load forecasting of regional power grids considering demand response under energy interconnection environment" (Li Chuang, Kong Xiangyu, Zhu Shijian, et al. Short-term load forecasting of regional power grids considering demand response under energy interconnection environment [J]. Automation of Electric Power Systems, 2021, 45(01): 71-78.) proposes a short-term load forecasting method that considers the impact of demand response electricity prices on load forecasting accuracy and improves forecasting accuracy. However, existing research mainly focuses on predicting electric loads based on electricity demand response, while there are few studies on load forecasting that consider the coupling relationship between multiple loads and demand response. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides a comprehensive energy system load forecasting method based on multi-energy demand response and ICPO-BiLSTM model. The forecasting method can effectively deal with the influence of the coupling relationship between multiple loads and user demand response on the load forecasting accuracy, and can improve the accuracy of multiple load forecasting.

[0007] The technical solution adopted by the present invention is:

[0008] The integrated energy system load forecasting method based on multi-energy demand response and ICPO-BiLSTM model includes the following steps:

[0009] Step 1: Based on the multi-energy demand response mechanism, extract the signal characteristics of the multi-energy demand response and build a multi-energy demand response model;

[0010] Step 2: Based on the chaos mapping theory and elite reverse learning strategy, the crown porcupine optimization algorithm CPO is improved;

[0011] Step 3: Optimize the BiLSTM neural network model based on the multi-head self-attention mechanism;

[0012] Step 4: Jointly forecast the multivariate loads.

[0013] In step 1, the amount of electricity transfer generated by the user's actual demand response can be expressed by the following formula:

[0014]

[0015] Where: T p , T f , T v They are peak, flat and valley periods of electricity prices respectively; is the actual electricity consumption transfer rate between peak and valley; are the fuzzy parameters of peak-to-flat and flat-to-valley electricity transfer rates respectively; They are the average loads at peak, normal and valley times before adopting time-of-use electricity prices; DR eis the electricity consumption shift caused by user demand response at time t.

[0016] When users participate in the time-of-use electricity price demand response, they will reduce part of the electricity load. At this time, the energy demand for the electricity load will turn to the cold and hot loads, and produce a coupled response with the cold and hot loads. It is necessary to calculate the demand response of the cold and hot loads to obtain the multi-energy demand response signal:

[0017]

[0018] CDR(Δp)=[DR e DR c DR h ];

[0019] Where: DR ch is the total response of cooling and heating loads; DR c ,DR h are the cold and hot load responses respectively; D is the response offset; η c , η h are the operating efficiency of cold and hot load equipment respectively; α ec , α eh are the response proportions of cold and hot loads respectively; Δp represents the peak-valley electricity price difference; CDR(Δp) represents the comprehensive demand response; [DR e DR c DR h ] represents the comprehensive demand response matrix.

[0020] The step 2 comprises:

[0021] 1) Initialize the population based on chaos mapping theory:

[0022] Since the initial population in the crested porcupine optimization algorithm CPO is randomly generated, there is an issue of uneven individual distribution during the solution process, and the global space cannot be effectively searched. The present invention uses the piecewise linear chaotic map (PWLCM) method to optimize the initialization process of the crested porcupine population to improve the spatial uniformity and population traversal of the initial population individuals, thereby enhancing the global search capability of the algorithm;

[0023]

[0024] Where: z(t+1) is the random position of the chaotic map; P Z is the control parameter of the mapping, P Z Take 0.4; z(t)∈[0,1] is the random iteration value generated, z(t)∈[0,1].

[0025] The initialization formula for optimizing the crested porcupine population based on chaos mapping theory can be expressed as follows:

[0026]

[0027] Where: is the i-th candidate solution after mapping; z(i) is the random value obtained by chaotic mapping; are the upper and lower bounds of the search space respectively; N′ is the number of individuals in the population; i represents the time step.

[0028] 2) Fusion elite reverse learning strategy:

[0029] On the basis of optimizing the initialization of the crested porcupine population, the present invention proposes an elite reverse learning strategy, which randomly selects a current optimal solution as the symmetry center and introduces random parameters so that the current solution corresponds to a reverse region. The reverse solution of each reverse learning is randomly generated from the corresponding region, thereby increasing the probability of successful reverse learning. The elite reverse solution generation method is as follows:

[0030]

[0031] Where: X c For the center of excellence; X o1 , X o2 , X o3 is the current optimal solution; is the reverse solution; X is the current solution; r 4 、r 5 、r 6 is a random number between (0,1).

[0032] Compare the fitness value of the elite reverse solution with the historical optimal fitness value of the retained solution. If the fitness value of the elite reverse solution is smaller than the historical optimal fitness value of the retained solution, the elite reverse solution is retained and there is no need to calculate the fitness of the current solution. If the fitness value of the elite reverse solution is larger than the historical optimal fitness value of the retained solution, it is necessary to calculate the fitness value of the current solution and retain the current solution. The update method of this strategy is as follows:

[0033]

[0034] Where: X * To retain the solution; is the fitness value of the elite reverse solution; fitness is the historical optimal fitness value.

[0035] In step 3, multiple correlations between input features are mined based on a multi-head self-attention mechanism, and the weights of each input feature are quantified to obtain an input feature weighted matrix X′, which is then input into a multi-layer BiLSTM neural network model to predict multivariate loads;

[0036] Based on the multi-head self-attention mechanism, various correlations between input features are mined, and the weights of each input feature are quantified to obtain the input feature weighted matrix X′. The expression of this process is as follows:

[0037]

[0038] α m =M H (Q m ,K m ,V m )=[α 1 ,α 2 ,…,α k ]

[0039] X′=diag(α 1 ,α 2 ,…α k )X

[0040] Where: Q m , K m 、V m They are query matrix, key matrix and value matrix respectively; They are used to calculate Q m , K m 、V m The parameter matrix of X is the input feature matrix; α m is the feature weight; M H (·) is the multi-head attention function; [Q m ,K m ,V m ] represents the joint matrix composed of the query matrix, key matrix and value matrix; Indicates the calculation of [Q m ,K m ,V m ] is the joint parameter matrix. 1 ,α 2 ,…,α k Respectively represent the weights of each input feature after quantization; [α 1 ,α 2 ,v,α k ] represents the weight matrix of the quantized input features; diag(α 1 ,α 2 ,…α k ) represents the diagonalization of the weights of each quantized input feature;

[0041] The specific structure of the multi-layer BiLSTM neural network model is as follows Figure 3As shown in Figure 1, the model consists of an information input layer, a result output layer, and two hidden layers in the middle;

[0042] Information input layer: input simulation training set (feature weighted matrix);

[0043] Result output layer: output the prediction results of multivariate loads;

[0044] Hidden layer 1: The first part is forward transmission, the simulation training set enters the input layer, and after being calculated and processed by the network model, the final result is submitted to the hidden layer 2. The second part is reverse propagation, and after being calculated and processed by the network model, the final result is submitted to the hidden layer 2;

[0045] Hidden layer 2: The result calculated in hidden layer 1 is used as the input value. The input value is forward transmitted through the first part to obtain the final result and reversely propagated through the second part to obtain the final result. Finally, the forward result and the reverse result are spliced ​​together to obtain the output result.

[0046] The BiLSTM model adds a reverse LSTM layer on the basis of the LSTM model to obtain a bidirectional neural network structure, thereby obtaining a better time series processing effect. Considering the complex relationship between electricity, heat, and cooling loads, as well as between multiple loads and weather factors, the present invention constructs a double hidden layer BiLSTM model to explore the deeper relationship among the influencing factors and improve the accuracy of multiple load prediction.

[0047] The expression form of the multi-layer BiLSTM network structure is:

[0048] a 2,2 =g(W a,2 [a 2,1 ,a 1,2 ]+b a,2 )

[0049]

[0050] Where: a i,j is the jth LSTM memory unit in the i-th layer; is the jth LSTM memory unit in the i-th layer; W a,i , W y are the weight matrices for the forward, reverse sequence and output of the i-th layer respectively; b a,i , b y are the forward and reverse sequences and the output bias vector of the i-th layer respectively; x, are input and output respectively.

[0051] In step 4, predicting the multivariate load includes the following steps:

[0052] 1) Set the size of the crested porcupine population and the maximum number of iterations.

[0053] 2) Generate the initialization population based on the piecewise linear chaos mapping theory.

[0054] 3) Based on the crown porcupine optimization algorithm ICPO, the global optimal solution is searched, and this optimal solution is used as the basis for parameter configuration of the multi-layer BiLSTM neural network model.

[0055] First, the initialization process of the crested porcupine population is improved based on the chaotic mapping theory, and the spatial uniformity and population traversal of the individual positions of the initial population are improved to enhance the global search ability of the ICPO algorithm. Then, based on the elite reverse learning strategy, the solution corresponding to the historical optimal fitness is obtained, which is the optimal parameter of the multi-layer BiLSTM neural network model.

[0056] 4) Integrate multivariate load historical data and external influencing factors to form an input feature matrix, and weight the input feature matrix through the MHSA module. By setting corresponding weights for different features, the input feature weighted matrix is ​​obtained.

[0057] The multi-load historical data refers to the historical usage of electricity, heat, and cooling loads. The research object is the comprehensive energy data set published by the Tempe campus of Arizona State University in the United States. The data set contains electricity, heat, and cooling load data from January to February 2024. The sampling interval is 1 hour, with a total of 168 sampling points. External influencing factors refer to the meteorological data of the corresponding time in the area where the campus is located, including temperature, air pressure, relative humidity, and wind speed.

[0058] Figure 4 The thermal diagram is obtained to calculate the maximum information coefficient between electricity, heating, cooling loads and multiple loads and meteorological factors. Figure 4 It can be seen that the MIC values ​​between the electric load and the cooling and heating loads are all greater than 0.83, indicating a strong coupling relationship. At the same time, the MIC value between the cooling load and the heating load is 0.8355, proving that there is also a strong correlation between the cooling and heating loads. In addition, the correlation between the electric, heating, and cooling loads and temperature, relative humidity, and wind speed is strong, and the correlation with air pressure is relatively weak.

[0059] In step 4), the MHSA (multi-head attention mechanism) module specifically includes:

[0060] Query matrix (Query, Q), key matrix (Key, K), value matrix (Value, V);

[0061] The query matrix is ​​a matrix used for comparison with the key matrix. They are extracted from the input sequence and used to explore other elements in the sequence that are related to the current element. The key matrix is ​​the index of the information stored in the sequence. They determine which positions in the sequence are important and are matched with the query matrix. The value matrix contains specific information at each position in the sequence. They are candidate information matched with the query matrix and are used for weighted summation after the attention score is calculated.

[0062] Scaling dot product attention: The result of the dot product is usually scaled to prevent the vanishing gradient problem. Scaling is achieved by dividing by a scaling factor (usually the square root of the key vector dimension).

[0063] Concat: Concatenate the outputs of all heads into a longer vector.

[0064] Linear transformation: The concatenated vectors are processed through a linear layer to integrate the information and transform it into the final output.

[0065] 5) Input the input feature weighted matrix obtained in step 4) into step 3), optimize the multi-layer BiLSTM neural network model and perform joint prediction on the multi-factor loads, and finally evaluate and verify the prediction results of the multi-factor loads.

[0066] The present invention provides a comprehensive energy system load forecasting method based on multi-energy demand response and improved BiLSTM, and the beneficial effects are as follows:

[0067] 1) In step 1 of the present invention, most of the existing literature only considers the application of power demand response in power load forecasting, while there is a relative lack of research on the coupled demand response that comprehensively considers different energy forms in multi-energy load forecasting. The present invention considers multi-energy demand response, integrates user demand response behavior, constructs input characteristic variables of multi-energy demand response, and more comprehensively analyzes the interactions and influences between different energy systems. By analyzing the coupled demand responses of different energy forms, the present invention can more accurately predict the changing trends of multiple loads, thereby improving the accuracy and reliability of load forecasting, and also helps to optimize energy scheduling strategies, achieve efficient allocation and use of energy, and reduce energy waste.

[0068] 2) In step 2 of the present invention, the chaos mapping theory helps to increase the algorithm's global search capability and ability to jump out of the local optimum due to its randomness and unpredictability, while the elite reverse learning strategy guides the search direction by back-propagating the error, thereby accelerating the algorithm's convergence to the optimal solution. The crowned porcupine optimization algorithm (CPO) is improved based on the chaos mapping theory and the elite reverse learning strategy. The improved CPO algorithm uses the randomness of the chaos mapping to break the convergence bottleneck that may be encountered in the traditional optimization process, and at the same time uses the elite reverse learning strategy to ensure the correctness of the search direction, thereby quickly locating a better solution in the parameter space of the neural network. The application of this algorithm provides the optimal parameter configuration for the neural network, thereby providing more accurate prediction results in multivariate load forecasting.

[0069] 3) In step 3 of the present invention, in order to solve the problem of missing key input feature information due to uneven weight distribution during model training, the present invention introduces a multi-head self-attention mechanism to assign appropriate weights to input features according to their importance, ensuring that the model can accurately identify and give appropriate attention to important features. The multi-head self-attention mechanism can capture multiple correlations in the multivariate load forecasting input data at the same time by performing multiple self-attention calculations in parallel and merging the results together, providing the model with a richer correlation representation. Optimizing the BiLSTM neural network model based on the multi-head self-attention mechanism can not only accelerate the training process of the model, but also further improve the load forecasting accuracy.

[0070] 4) In step 4 of the present invention, the joint prediction of multiple loads takes into account the mutual influence and coupling relationship between different loads, and can more accurately capture the trend of load changes, thereby improving the accuracy of the prediction results. In addition, the joint prediction can simultaneously predict the three loads of electricity, heat, and cooling, greatly reducing the calculation cost, and has higher training efficiency than a single prediction model. Accurate and efficient load forecasting helps to better integrate renewable energy such as wind power and solar energy, balance supply and demand, and improve the level of renewable energy consumption. At the same time, the joint prediction can better predict the impact of extreme weather and emergencies on the load, helping system operators to prepare in advance and reduce potential risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0072] Figure 1 It is a flow chart of the integrated energy system load forecasting method of the present invention.

[0073] Figure 2(a) is a comparison of the electric load prediction results of the integrated energy system based on different models;

[0074] Figure 2(b) is a comparison of the heat load prediction results of the integrated energy system based on different models;

[0075] Figure 2(c) is a comparison of the cooling load prediction results of the integrated energy system based on different models.

[0076] Figure 3 This is a structural diagram of the multi-layer BiLSTM neural network model.

[0077] Figure 4 Thermodynamic diagram obtained by calculating the maximum information coefficient between electricity, heating, cooling loads and between multiple loads and meteorological factors.

[0078] Figure 5 Schematic diagram of weighting the input feature matrix for the MHSA module. DETAILED DESCRIPTION

[0079] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0080] Figure 1 This is a flow chart of the load forecasting method for an integrated energy system based on multi-energy demand response and improved BiLSTM. First, based on the multi-energy demand response mechanism, the signal characteristics of the multi-energy demand response are extracted, and a multi-energy demand response model is constructed; secondly, the crown porcupine optimization algorithm is improved based on the chaos mapping theory and the elite reverse learning strategy; then, the neural network of BiLSTM is optimized based on the multi-head self-attention mechanism. Finally, the load forecasting method for an integrated energy system based on multi-energy demand response and improved BiLSTM is used to predict multiple loads. The prediction method proposed in the present invention can effectively deal with the influence of the coupling relationship between multiple loads and the user demand response on the load prediction accuracy, and can improve the accuracy of the prediction of multiple loads.

[0081] Figure 2(a) and Figure 2(b) are comparison diagrams of multivariate load forecasting results of integrated energy systems based on different models. To verify the superiority and reliability of the ICPO-BiLSTM-MHSA model proposed in this invention;

[0082] The LSTM model, BiLSTM model and ICPO-BiLSTM model are used as comparison models, and the prediction results of the ICPO-BiLSTM-MHSA model proposed in the present invention are compared and analyzed.

[0083] As can be seen from the electric load prediction result diagram in Figure 2 (a), during the period of 29-34 hours, the electric load curve predicted by the model proposed in the present invention is most consistent with the actual load curve. This is because the model uses the ICPO algorithm to optimize the hyperparameters of BiLSTM, providing the optimal parameter configuration for the BiLSTM neural network, thereby improving the accuracy of the prediction.

[0084] It can be seen from the heat load prediction result diagram of Figure 2(b) that the heat load prediction result curve based on the model proposed in the present invention is closest to the heat load true value curve and has the highest prediction accuracy. This is because the heat load demand is greatly affected by meteorological factors. The present invention assigns weights to different input features based on MHSA, so that the model can give corresponding attention to different input features according to the set weights, thereby improving the prediction accuracy of the heat load.

[0085] As can be seen from the cooling load prediction result diagram in Figure 2(c), the cooling load fluctuation curve has a relatively obvious regularity. During the period of 108-116h, the model proposed in the present invention can better identify the increasing trend of cooling load in the system and achieve a good prediction effect. However, the models based on LSTM and BiLSTM failed to respond to the load changes in time, and the prediction accuracy was not high.

[0086] Table 1 Comparison of prediction errors with and without multi-energy demand response

[0087]

[0088]

[0089] Table 1 is a comparison of the prediction errors with and without considering multi-energy demand response. As can be seen from Table 1, the prediction error obtained by taking into account multi-energy demand response is smaller than that obtained without taking into account multi-energy demand response. The evaluation index RMSE of electricity, heat and cooling load prediction taking into account multi-energy demand response is reduced by 71.38KW, 0.07mmBTU and 48kt respectively; the evaluation index MAPE% is reduced by 6.59%, 13.04% and 24.86% respectively. This is because considering multi-energy demand response in multi-element load forecasting can help the model understand the operating characteristics of the integrated energy system more comprehensively and improve its prediction ability.

Claims

1. A comprehensive energy system load forecasting method based on multi-energy demand response and ICPO-BiLSTM model, characterized by The following steps are involved: Step 1: Based on the multi-energy demand response mechanism, extract the signal characteristics of the multi-energy demand response and build a multi-energy demand response model; Step 2: Based on the chaos mapping theory and elite reverse learning strategy, the crown porcupine optimization algorithm CPO is improved; Step 3: Optimize the BiLSTM neural network model based on the multi-head self-attention mechanism; Step 4: Jointly forecast the multivariate loads; In step 1, the amount of electricity transfer generated by the user's actual demand response is expressed by the following formula: Where: T p 、T f 、T v They are peak, flat and valley periods of electricity prices respectively; is the actual electricity consumption transfer rate between peak and valley; are the fuzzy parameters of peak-to-flat and flat-to-valley electricity transfer rates respectively; are the average loads at peak, normal and valley times before adopting time-of-use electricity prices; DR e is the electricity consumption shift caused by user demand response at time t; When users participate in the time-of-use electricity price demand response, they will reduce part of the electricity load. At this time, the energy demand for the electricity load will turn to the cold and hot loads, and produce a coupled response with the cold and hot loads; calculate the demand response of the cold and hot loads to obtain the multi-energy demand response signal: CDR(Δp)=[DR e DR c DR h ]; Where: DR ch is the total response of cooling and heating loads; DR c ,DR h are the cold and hot load responses respectively; D is the response offset; η c , η h are the operating efficiency of cold and hot load equipment respectively; α ec , α eh are the response proportions of cold and hot loads respectively; Δp represents the difference between peak and valley electricity prices; CDR(Δp) represents comprehensive demand response; [DR e DR c DR h ] represents the comprehensive demand response matrix; The step 2 comprises: 1) Initialize the population based on chaos mapping theory: The piecewise linear chaotic mapping method is used to optimize the initialization process of the crested porcupine population to improve the spatial uniformity and population traversal of the initial population individuals, thereby enhancing the global search ability of the algorithm. Where: z(t+1) is the random position of the chaotic map; P Z is the control parameter of the mapping; z(t) is the random iteration value generated, z(t)∈[0,1]; The initialization formula for optimizing the crested porcupine population based on chaos mapping theory is expressed as follows: Where: is the i-th candidate solution after mapping; z(i) is the random value obtained by chaotic mapping; are the upper and lower bounds of the search space respectively; N′ is the number of individuals in the population; i represents the time step; 2) Fusion elite reverse learning strategy: On the basis of optimizing the initialization of the crested porcupine population, an elite reverse learning strategy is proposed. This strategy randomly selects a current optimal solution as the symmetry center and introduces random parameters to make the current solution correspond to a reverse region. The reverse solution of each reverse learning is randomly generated from the corresponding region, thereby increasing the probability of successful reverse learning. The elite reverse solution generation method is as follows: Where: X c For the center of excellence; X o1 , X o2 , X o3 is the current optimal solution; is the reverse solution; X is the current solution; r4, r5, r6 are random numbers between (0,1); Compare the fitness value of the elite reverse solution with the historical optimal fitness value of the retained solution. If the fitness value of the elite reverse solution is smaller than the historical optimal fitness value of the retained solution, the elite reverse solution is retained and the fitness of the current solution does not need to be calculated; if the fitness value of the elite reverse solution is larger than the historical optimal fitness value of the retained solution, the fitness value of the current solution needs to be calculated and the current solution is retained. The update method of this strategy is as follows: Where: X * To retain the solution; is the fitness value of the elite reverse solution; fitness is the historical optimal fitness value; in step 3, multiple correlations between input features are mined based on the multi-head self-attention mechanism, and the weights of each input feature are quantified to obtain the input feature weighted matrix X′, which is then input into the multi-layer BiLSTM neural network model to predict the multivariate load; Based on the multi-head self-attention mechanism, various correlations between input features are mined, and the weights of each input feature are quantified to obtain the input feature weighted matrix X′. The expression of this process is as follows a m =M H (Q m ,K m ,V m )=[α1,α2,…,α k ] X′=diag(α1,α2,…α k )X Where: Q m , K m 、V m They are query matrix, key matrix and value matrix respectively; They are used to calculate Q m , K m 、V m The parameter matrix of X is the input feature matrix; α m is the feature weight; M H (·) is the multi-head attention function; [Q m ,K m ,V m ] represents the joint matrix composed of the query matrix, key matrix and value matrix; Indicates the calculation of [Q m ,K m ,V m ] joint parameter matrix; α1,α2,…,α k Respectively represent the weights of each input feature after quantization; [α1, α2, …, α k ] represents the weight matrix of the quantized input features; diag(α1,α2,…α k ) represents the diagonalization of the weights of each quantized input feature; The BiLSTM model adds a reverse LSTM layer to the LSTM model to obtain a bidirectional neural network structure. The expression form of the multi-layer BiLSTM network structure is: a 2,2 =g(W a,2 [a 2,1 ,a 1,2 ]+b a,2 ) Where: a i,j is the jth LSTM memory unit in the i-th layer; is the jth LSTM memory unit in the i-th layer; W a,i , W y are the weight matrices for the forward, reverse sequence and output of the i-th layer respectively; b a,i , b y are the bias vectors of the forward and reverse sequences and output of the i-th layer respectively; In step 4, predicting the multivariate load includes the following steps: 1) Set the size of the crested porcupine population and the maximum number of iterations; 2) Generate the initialization population based on the piecewise linear chaos mapping theory; 3) Search for the global optimal solution based on the crown porcupine optimization algorithm ICPO, and use this optimal solution as the basis for parameter configuration of the multi-layer BiLSTM neural network model; 4) Integrate multivariate load historical data and external influencing factors to form an input feature matrix, and weight the input feature matrix through the MHSA module. By setting corresponding weights for different features, the input feature weighted matrix is ​​obtained; 5) Input the input feature weighted matrix obtained in step 4) into step 3), optimize the multi-layer BiLSTM neural network model and perform joint prediction on multivariate loads.

2. The integrated energy system load forecasting method based on multi-energy demand response and ICPO-BiLSTM model according to claim 1 is characterized by: Step 4) includes: query matrix, key matrix, value matrix; The query matrix is ​​a matrix used for comparison with the key matrix. They are extracted from the input sequence to explore other elements in the sequence that are related to the current element. The key matrix is ​​an index of the information stored in the sequence, which determines which positions of information in the sequence are important and is matched with the query matrix; The value matrix contains the specific information of each position in the sequence. They are the candidate information matching the query matrix and are used for weighted summation after the attention score is calculated. Scaled dot product attention: The result of the dot product is scaled to prevent the gradient vanishing problem; scaling is achieved by dividing by a scaling factor; Concatenate: concatenate the outputs of all heads into a longer vector; Linear transformation: The concatenated vectors are processed through a linear layer to integrate the information and transform it into the final output.

Images