Comprehensive energy system short-term multi-element load prediction method fusing deep learning combination model

Abstract

The invention discloses an integrated energy system short-term multivariate load prediction method fusing a deep learning combination model, and the method comprises the steps: obtaining the multivariate load historical data and meteorological data of an integrated energy system, and obtaining a correlation factor which is greatly affected by the multivariate load; dividing the multi-element load historical data into two parts, namely a training set and a test set, and performing maximum and minimum normalization processing; extracting multi-level spatial features of the multi-element load historical data by using a series-parallel superposed multi-scale convolutional neural network; load historical data time sequence features are extracted by using a bidirectional gating cycle unit neural network, and a multi-head attention mechanism is introduced to extract important information of a load prediction task, so that a multi-scale convolution bidirectional gating cycle unit neural network fused with the attention mechanism is designed; and a self-adaptive multi-strategy grey wolf optimization algorithm is utilized to realize optimization parameter search of the bidirectional gating circulation unit neural network, an evaluation index is adopted to carry out error analysis on a prediction result, an optimal parameter is obtained, and accurate load prediction is realized.

Description

Technical Field

[0001] The present invention relates to a short-term multivariate load forecasting method for an integrated energy system integrating a deep learning combined model. Background Art

[0002] In the context of energy internet and low-carbon life, the integrated energy system has gotten rid of the independence of various energy sources in the design and operation of the traditional energy system, and has realized the integration of production, supply and marketing from energy generation to consumption, and achieved the purpose of mutual transformation and unified management of multiple energy networks through energy conversion equipment. Among them, accurate and efficient load forecasting is an important prerequisite for the operation planning and optimal scheduling of the entire integrated energy system. The integrated energy system involves the production, transmission, storage, conversion and utilization of various energy forms such as cold, heat and electricity. While fully absorbing renewable energy and improving energy utilization, it realizes the complementary and coordinated operation of multiple energy sources, and ultimately achieves the goal of environmental friendliness and sustainable development. Load forecasting is the basis of energy management and optimal scheduling of the integrated energy system, and its prediction accuracy is directly related to the overall operation performance of the system. However, a large amount of disorderly energy use will inevitably have an impact on the economic operation and scheduling of the energy network. Therefore, improving the accuracy of multi-load forecasting is of great significance to guiding the orderly energy use of the integrated energy system.

[0003] Currently commonly used load forecasting methods include traditional forecasting methods and machine learning methods.

[0004] Traditional prediction methods mainly include exponential smoothing, trend extrapolation, time series and regression analysis. These prediction methods are only applicable to systems that are stable in the past, present and future, and cannot cope with random problems. Machine learning prediction methods mainly include support vector machines, neural networks, random forests and other methods. Among them, long short-term memory networks are a machine learning prediction method that has been widely used in recent years, and the bidirectional long short-term memory neural network based on its extension has also been widely used. Long short-term memory neural networks can effectively solve problems such as falling into local optimal solutions and gradient explosion, and they perform better in time series data training. However, the above models are all single models, and the results are not good in the prediction of multi-load in complex integrated energy systems. In recent years, the introduction of combined models has effectively improved the accuracy of multi-load prediction, such as convolution-bidirectional long short-term memory networks and convolution-bidirectional gated recurrent units. When the combined model network is applied to the comprehensive prediction of electric load, heat load and cooling load in the field of integrated energy, the coupling of multi-loads puts higher requirements on the prediction performance. Summary of the invention

[0005] The present invention aims to solve the problems existing in the above-mentioned prior art and provides a short-term multi-element load forecasting method for an integrated energy system integrating a deep learning combined model.

[0006] The technical solutions adopted in the present invention are:

[0007] A short-term multivariate load forecasting method for an integrated energy system integrating a deep learning combined model comprises the following steps:

[0008] S1: Obtain the multi-load historical data and meteorological data of the integrated energy system, optimize the correlation factors of the multi-load historical data and meteorological data, and obtain the correlation factors with greater impact on the multi-load;

[0009] S2: Divide the multivariate load historical data into a training set and a test set, and perform maximum and minimum normalization on the training set and the test set;

[0010] S3: Extracting multi-level spatial features of multivariate load history data using a multi-scale convolutional neural network with serial and parallel superposition;

[0011] The bidirectional gated recurrent unit neural network is used to extract the time series characteristics of the load history data, and a multi-head attention mechanism is introduced to extract important information of the load forecasting task, so as to design a multi-scale convolutional bidirectional gated recurrent unit neural network integrating the attention mechanism;

[0012] S4: The adaptive multi-strategy grey wolf optimization algorithm is used to implement the bidirectional gated recurrent unit neural network to search for optimization parameters, and the mean absolute error, root mean square error and mean relative error evaluation indicators are used to analyze the error of the prediction results to obtain the optimal parameters and achieve accurate load prediction.

[0013] Furthermore, in S1:

[0014] Assume that X represents meteorological data, Y represents multivariate load historical data, and use the Pearson correlation coefficient to optimize the correlation factors of multivariate load historical data and meteorological data to obtain the meteorological factors with greater impact on multivariate loads. The expression of the Pearson correlation coefficient used is as follows:

[0015]

[0016] Where: x i represents the meteorological data at the i-th moment, y i represents the multivariate load historical data at the i-th moment, n represents the number of samples, cov(X,Y) represents the covariance of X and Y, σ X and σ Y represent the standard deviation of X and Y respectively.

[0017] Furthermore, in S2, the calculation formula for the maximum and minimum normalization processing is:

[0018]

[0019] In the formula: x represents the original sample data, x min Indicates the minimum sample value; x max Indicates the maximum sample value; Represents the sample value after normalization.

[0020] Furthermore, the serial-parallel superposition multi-scale convolutional neural network includes two convolutional neural networks, and the size and number of convolution kernels of the first convolutional layer and the second convolutional layer in each convolutional neural network are determined according to task requirements and data dimensions. The convolution calculation formula of the convolutional neural network is:

[0021] c i =T([x i :x i+f-1 ]*K i +b i )

[0022] C=[c1,c2,…,c i+f-1 ]

[0023] Among them, K represents the convolution kernel, i represents the width of the convolution kernel, f represents the convolution scale, T is the activation function, and b i is the bias, x i :x i+f-1 Represents the matrix vectors from i to i+f-1 in X.

[0024] 5. The short-term multivariate load forecasting method for an integrated energy system integrating a deep learning combined model as described in claim 1 is characterized by: using a bidirectional gated recurrent unit neural network to learn the time series characteristics of load data and iteratively extract global information; the bidirectional gated recurrent unit neural network is provided with two layers of bidirectional gated recurrent units, and the number of hidden layer neurons in each layer of bidirectional gated recurrent units is determined by an adaptive multi-strategy grey wolf optimization algorithm, and a Dropout layer is introduced into the hidden layer.

[0025] Furthermore, the calculation expression of the gated recurrent unit is as follows:

[0026] z t =σ(W iz x t +b iz +W hz h t-1 +b ht )

[0027] r t =σ(W ir x t +b ir +W hr h t-1 +b hr )

[0028] h t =tanh[W in x t +b in +r t ⊙(W hn h t-1 +b hn )]

[0029] h t =(1-z t )⊙h t-1 +z t ⊙h t

[0030] Where: h t-1 is the hidden state at time t-1; x t is the input at time t; σ is the sigmoid activation function; r t is the reset gate; t is the update gate; h t is the final hidden state; W is the coefficient matrix; ⊙ is the Hadamard product; b is the paranoid coefficient;

[0031] The bidirectional gated recurrent unit is composed of two-way gated recurrent units. The time series data is calculated in the two gated recurrent units in opposite directions, which are responsible for processing historical information and future information respectively. For an input sequence X = x1, x2, ..., x T , the forward computation of the bidirectional gated recurrent unit is expressed as:

[0032]

[0033] Where: and are the hidden states from left to right and from right to left, respectively, x t is the tth element of the input sequence; on this basis, the hidden states in the two directions are fused to obtain the final hidden state:

[0034]

[0035] Pass the hidden state to a fully connected layer to get the output y t :

[0036] y t =softmax(Wh t +b).

[0037] Furthermore, the attention mechanism is a multi-head attention mechanism, and the expression is:

[0038]

[0039] MultiHead(Q,K,V)=concat(head1,…,head h )

[0040] head i =Attention(QW i Q ,KW i K ,VW i V )

[0041] Where: Q is the query vector; K is the key vector, V is the value vector; softmax is the weight normalization function; head i represents the i-th attention head; h is the number of heads in the multi-head self-attention mechanism; W i Q , W i K , W i V are the learnable weight matrices corresponding to Q, K, and V in the i-th attention; d k is the dimension of the K matrix; QK T QW is the product of the query matrix and the transposed key matrix; i Q KW is the product of the query matrix and the corresponding weight matrix in the i-th attention head; i K VW is the product of the key matrix and the corresponding weight matrix in the i-th attention head; i V is the product of the median matrix of the i-th attention head and the corresponding weight matrix; Attention(Q,K,V) is the output attention value; MultiHead(Q,K,V) is the output multi-head attention value; Attention(QW i Q , KW i K , VW i V ) is the output of the i-th attention value.

[0042] Furthermore, the adaptive multi-strategy gray wolf optimization algorithm is an improvement on the traditional gray wolf optimization algorithm, and the improvements are as follows:

[0043] (1) Initialization population of chaos model:

[0044] The infinite folding SIN chaos model is used to initialize the population, and its expression is:

[0045]

[0046] (2) Nonlinear convergence factor:

[0047] A nonlinear convergence factor is used to adapt and optimize the chaotic model. The expression of the nonlinear convergence factor is:

[0048]

[0049] Where: a init with a fin are the initial values ​​of the convergence factors respectively;

[0050] (3) Dimensional learning individual update:

[0051] The position update method based on dimensional learning between wolf individuals is introduced to improve the search ability of population individuals. The specific strategy is:

[0052] First, the position X of each individual gray wolf i is calculated using the Euclidean distance i (t) and X i～G (t+1) domain range R i (t) is:

[0053] R i (t)=||X i (t)-X i～G (t+1)||

[0054] Secondly, according to the R of each individual gray wolf i (t), construct the neighboring gray wolf group N of each individual gray wolf i (t) is:

[0055]

[0056] On this basis, multi-neighbor learning is performed, which is expressed as:

[0057] X i～D (t+1)=X i (t)+rand·(X n (t)-X r (t))

[0058] Finally, the new position of the individual is updated according to the fitness value, which is expressed as:

[0059]

[0060] (4) Generalized reverse learning variation:

[0061] A generalized reverse learning strategy is used to perturb the position of the leader wolf at the end of each search, which is expressed as:

[0062]

[0063] Where, X i a / b / d is the j-th dimension of the position of the leading wolves a, b and d, and k is the proportional factor; after the leading wolves a, b and d are disturbed, the fitness values ​​before and after the disturbance are compared to determine whether the leading wolves a, b and d update their positions.

[0064] Furthermore, the evaluation index is: root mean square error E RMSE , mean absolute error E MAE And the average relative error E MAPE , the specific calculation is as follows:

[0065]

[0066] Where: n represents the number of predicted loads, y i represents the actual value of the i-th load, represents the predicted value of the i-th load.

[0067] The present invention has the following beneficial effects:

[0068] 1) The method of the present invention integrates load characteristics and meteorological factors to construct a multidimensional characteristic information matrix, which provides data support for the multivariate load forecasting model and is conducive to improving the accuracy of load forecasting.

[0069] 2) The method of the present invention utilizes serial-parallel superposition of multi-scale convolutional neural networks and bidirectional gated recurrent unit neural networks to simultaneously mine the spatiotemporal characteristics of load and meteorological sequences, and introduces a multi-head attention mechanism to extract important information for load forecasting tasks, and establishes a multivariate load forecasting model of a multi-scale convolutional bidirectional gated recurrent unit neural network integrating an attention mechanism.

[0070] 3) The method of the present invention utilizes an adaptive multi-strategy grey wolf optimization algorithm to optimize the number of hidden layer neurons of the bidirectional gated recurrent unit neural network model, the learning rate and the number of iterations of the combined model, thereby improving the search efficiency, solution accuracy and convergence speed, thereby achieving better performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 This is the structure diagram of the multi-scale convolutional bidirectional gated recurrent unit neural network model that integrates the attention mechanism.

[0072] Figure 2 It is a flow chart of the present invention.

[0073] Figure 3 It is a comparison diagram of the predicted value and actual value of the electric load of the present invention.

[0074] Figure 4 It is a comparison chart of the heat load prediction value and the actual value of the present invention. DETAILED DESCRIPTION

[0075] The present invention will be further described below in conjunction with the accompanying drawings.

[0076] like Figure 1 and Figure 2 The present invention provides a short-term multivariate load forecasting method for an integrated energy system integrating a deep learning combined model, and the steps are as follows:

[0077] S1: Obtain the multi-load historical data and meteorological data of the integrated energy system, optimize the correlation factors of the multi-load historical data and meteorological data, and obtain the correlation factors with greater impact on the multi-load;

[0078] S2: further divide the multivariate load historical data into a training set and a test set, and perform maximum and minimum normalization processing on the training set and the test set;

[0079] S3: Extract multi-level spatial features of load history data using a multi-scale convolutional neural network with serial and parallel superposition;

[0080] The bidirectional gated recurrent unit neural network is used to extract the time series characteristics of the load history data, and a multi-head attention mechanism is introduced to extract important information of the load forecasting task, so as to design a multi-scale convolutional bidirectional gated recurrent unit neural network integrating the attention mechanism;

[0081] S4: The adaptive multi-strategy grey wolf optimization algorithm is used to implement the bidirectional gated recurrent unit network for optimization parameter search, and the mean absolute error, root mean square error and mean relative error evaluation indicators are used to perform error analysis on the prediction results to obtain the optimal parameters and achieve accurate load prediction.

[0082] The prediction method of the present invention integrates the respective advantages of deep learning models, which not only improves the accuracy of short-term multi-element load forecasting, but also facilitates the economic dispatch and efficient operation of the integrated energy system.

[0083] In S1, the load and weather data of a comprehensive energy system from January 1, 2020 to April 1, 2020 are selected. The time interval for recording data is 1 hour, with a total of 2184 sets of sample data. The weather data includes dew point, surface reflectivity, rainfall, temperature, air pressure and solar zenith angle.

[0084] In S1: Pearson correlation coefficient is used to optimize the correlation factors of multi-load historical data and meteorological data, and the meteorological factors with greater impact on multi-load are screened out. The specific process is as follows:

[0085] Assume that X represents meteorological data and Y represents multivariate load historical data. The expression of Pearson correlation coefficient is as follows:

[0086]

[0087] Where: x i represents the meteorological data at the i-th moment; y i represents the multivariate load historical data at the i-th moment; n represents the number of samples; cov(X,Y) represents the covariance of X and Y; σ X and σ Y represent the standard deviation of X and Y respectively.

[0088] The Pearson correlation coefficient matrix shows that the Pearson correlation coefficients of the six weather factors, namely dew point, surface reflectivity, rainfall, temperature, air pressure and solar zenith angle, are all greater than 0.6, namely 0.652, 0.587, 0.673, 0.732, 0.621 and 0.567 respectively. Therefore, these six are selected as the weather factor input data of the multivariate load.

[0089] In S2: the first 80% of the 2184 groups of sample data are used as the training set, and the last 20% are used as the test set. The training set and the test set are subjected to maximum and minimum normalization. The specific calculation is as follows:

[0090]

[0091] Where: x represents the original sample data, x min Indicates the minimum sample value; x max Indicates the maximum sample value; Represents the sample value after normalization.

[0092] In S3: the multi-scale convolutional neural network with serial-parallel superposition is composed of two convolutional neural networks with different convolution kernels in parallel, where the activation function is sigmoid, the pooling layer is one layer, and average pooling is adopted.

[0093] The convolution kernel of the first convolution layer is 1×1, the number is 32, and the convolution kernel of the second convolution layer is 1×1, the number is 64. The convolution calculation is as follows:

[0094] c i =T([x i :x i+f-1 ]*K i +b i )

[0095] C=[c1,c2,…,c i+f-1 ]

[0096] Wherein, K represents the convolution kernel, i represents the width of the convolution kernel, f represents the convolution scale, and T is the activation function; b i is the bias; x i :x i+f-1 Represents the matrix vectors from i to i+f-1 in X.

[0097] In S3: a bidirectional gated recurrent unit neural network is used to learn the time series characteristics of the load data and iteratively extract the global information.

[0098] The bidirectional gated recurrent unit neural network sets two layers of bidirectional gated recurrent units, uses the LeakyReLU function as the activation function, and sets the number of neurons in the two hidden layers to range from 50 to 100. A dropout layer is introduced after each hidden layer of the bidirectional gated recurrent unit, and the weight loss coefficient of the Dropout layer is set to 0.25, which can prevent overfitting, improve model generalization, and reduce model training time.

[0099] The calculation expression of the gated recurrent unit is as follows:

[0100] z t =σ(W iz x t +b iz +W hz h t-1 +b ht )

[0101] r t =σ(W ir x t +b ir +W hr h t-1 +b hr )

[0102] h t =tanh[W in x t +b in +r t ⊙(W hn h t-1 +b hn )]

[0103] h t =(1-z t )⊙h t-1 +z t ⊙h t

[0104] Where: h t-1 is the hidden state at time t-1; x t is the input at time t; σ is the sigmoid activation function; r t is the reset gate; t is the update gate; h t is the final hidden state; W is the coefficient matrix; ⊙ is the Hadamard product; b is the paranoid coefficient.

[0105] The bidirectional gated recurrent unit is composed of gated recurrent units in two directions. The time series data is calculated in the two gated recurrent units in opposite directions, which are responsible for processing historical information and future information respectively. For an input sequence X = x1, x2, ..., x T , the forward computation of the bidirectional gated recurrent unit can be expressed as:

[0106]

[0107] Where: and are the hidden states from left to right and from right to left, respectively, x t is the tth element of the input sequence.

[0108] On this basis, the hidden states in the two directions are fused to obtain the final hidden state:

[0109]

[0110] Pass the hidden state to a fully connected layer to get the output y t :

[0111] y t =softmax(Wh t +b)

[0112] In S3: The attention mechanism layer in the design adopts a multi-head attention mechanism, in which 3 heads are designed, and the number of heads is set to 2, 4 and 8 respectively, the number of channels of the attention mechanism is set to 16, and the weight coefficient is set to 0.1. It can simultaneously consider the impact of different features on the output and obtain the output matrix. The self-attention mechanism is calculated as follows:

[0113]

[0114] MultiHead(Q,K,V)=concat(head1,…,head h )

[0115] head i =Attention(QW i Q ,KW i K ,VW i V )

[0116] Where: Q is the query vector; K is the key vector, V is the value vector; softmax is the weight normalization function; head i represents the i-th attention head; h is the number of heads in the multi-head self-attention mechanism; W i Q , Wi K , W i V are the learnable weight matrices corresponding to Q, K, and V in the i-th attention; d k is the dimension of the K matrix; QK T QW is the product of the query matrix and the transposed key matrix; i Q KW is the product of the query matrix and the corresponding weight matrix in the i-th attention head; i K VW is the product of the key matrix and the corresponding weight matrix in the i-th attention head; i V is the product of the median matrix of the i-th attention head and the corresponding weight matrix; Attention(Q,K,V) is the output attention value; MultiHead(Q,K,V) is the output multi-head attention value; Attention(QW i Q , KW i K , VW i V ) is the output of the i-th attention value.

[0117] In S4: Adaptive multi-strategy gray wolf optimization algorithm is used to optimize the hyperparameters of the bidirectional gated recurrent unit (such as the number of hidden layer neurons, learning rate, and number of iterations), wherein the learning rate range is set to [0.001, 0.01], the number of iterations is set to [50, 200], the learning rate reduction factor is set to 0.4, the number of batch samples is 30, and other model parameters are optimized using adaptive moment estimation optimization algorithm and layer normalization technology to train a multi-scale convolutional bidirectional gated recurrent unit neural network model integrating attention mechanism. The traditional gray wolf optimization algorithm is prone to fall into the local optimal problem. The present invention proposes an adaptive multi-strategy gray wolf optimization algorithm, which is an improvement on the traditional gray wolf optimization algorithm, and the improvements are as follows:

[0118] The specific improvement strategies are as follows:

[0119] (1) Initialization population of chaos model:

[0120] In order to improve the diversity and distribution quality of the initial population, the infinite folding SIN chaos model with more chaotic characteristics than other chaos models is used to initialize the population. Its expression is:

[0121]

[0122] (2) Nonlinear convergence factor:

[0123] In order to balance the global and local search capabilities of the network model, a nonlinear convergence factor is used to adapt the optimization model. The expression of the nonlinear convergence factor is:

[0124]

[0125] Where: a init with a fin are the initial values ​​of the convergence factors.

[0126] (3) Dimensional learning individual update:

[0127] In view of the problem that the wolf pack individuals in the gray wolf algorithm only update and iterate according to the position of the wolf, thus ignoring the communication information between the wolf pack individuals, a position update method based on dimensional learning between wolf pack individuals is introduced to further improve the search ability of the population individuals. The specific strategy is:

[0128] First, the position X of each individual gray wolf i is calculated using the Euclidean distance i (t) and X i～G (t+1) domain range R i (t) is:

[0129] R i (t)=||X i (t)-X i～G (t+1)||

[0130] Secondly, according to the R of each individual gray wolf i (t), construct the neighboring gray wolf group N of each individual gray wolf i (t) is:

[0131]

[0132] On this basis, multi-neighbor learning is performed, which is expressed as:

[0133] X i～D (t+1)=X i (t)+rand·(X n (t)-X r (t))

[0134] Finally, the new position of the individual is updated according to the fitness value (minimization problem), which is expressed as:

[0135]

[0136] (4) Generalized reverse learning variation:

[0137] In order to enhance the ability of the algorithm to jump out of the local optimum in the later stage of the search, a generalized reverse learning strategy is used to perturb the position of the leader wolf in the later stage of each search. When trapped in the local optimum, it can search in the opposite direction with a large step length to enhance the probability of jumping out of the local trap. It is expressed as:

[0138]

[0139] Where, X i a / b / d is the j-th dimension of the position of the leading wolves a, b and d, and k is the proportional factor; after the leading wolves a, b and d are disturbed, the fitness values ​​before and after the disturbance are compared to determine whether the leading wolves a, b and d update their positions.

[0140] The three evaluation indicators of the present invention are specifically calculated as follows:

[0141]

[0142] Where: n represents the number of predicted loads; y i represents the actual value of the i-th load; represents the predicted value of the i-th load.

[0143] In order to fully demonstrate the effectiveness of the deep learning combination model and the adaptive multi-strategy optimization algorithm proposed in this invention, an experiment on the prediction of electric and thermal loads in a comprehensive energy system was designed and evaluated. The electric and thermal loads were predicted using the multi-scale convolutional bidirectional gated recurrent unit neural network model with attention mechanism proposed in this patent. The prediction results are as follows: Figure 3 and Figure 4 As shown. The calculated error is used to obtain the average absolute error E of the electric load. MAE is 15.2745, the root mean square error E RMSE The average relative error is 15.24 and E MAPE The average absolute error of heat load is 1.5136%. MAE is 20.2786, the root mean square error E RMSE The average relative error is 30.2733 and E MAPE The above results show that the model constructed by the present invention can achieve higher prediction accuracy, which confirms the superiority and effectiveness of the present invention.

[0144] The above description is only a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be regarded as within the protection scope of the present invention.

Claims

1. A short-term multivariate load forecasting method for an integrated energy system integrating a deep learning combined model, characterized by: The following steps are involved: S1: Obtain the multi-load historical data and meteorological data of the integrated energy system, optimize the correlation factors of the multi-load historical data and meteorological data, and obtain the correlation factors with greater impact on the multi-load; S2: Divide the multivariate load historical data into a training set and a test set, and perform maximum and minimum normalization on the training set and the test set; S3: Extract multi-level spatial features of multivariate load history data using a multi-scale convolutional neural network with serial and parallel superposition; The bidirectional gated recurrent unit neural network is used to extract the time series characteristics of the load history data, and a multi-head attention mechanism is introduced to extract important information of the load forecasting task, so as to design a multi-scale convolutional bidirectional gated recurrent unit neural network integrating the attention mechanism; S4: The adaptive multi-strategy grey wolf optimization algorithm is used to implement the bidirectional gated recurrent unit neural network to search for optimization parameters, and the evaluation index is used to perform error analysis on the prediction results to obtain the optimal parameters and realize accurate load prediction.

2. The method for short-term multivariate load forecasting of an integrated energy system integrating a deep learning combined model as claimed in claim 1, characterized in that: In S1: Assume that X represents meteorological data, Y represents multivariate load historical data, and use the Pearson correlation coefficient to optimize the correlation factors of multivariate load historical data and meteorological data to obtain the meteorological factors with greater impact on multivariate loads. The expression of the Pearson correlation coefficient used is as follows: Where: x i represents the meteorological data at the i-th moment, y i represents the multivariate load historical data at the i-th moment, n represents the number of samples, cov(X,Y) represents the covariance of X and Y, σ X and σ Y represent the standard deviation of X and Y respectively.

3. The short-term multivariate load forecasting method for an integrated energy system integrating a deep learning combined model as claimed in claim 1, characterized in that: In S2, the calculation formula for the maximum and minimum normalization processing is: In the formula: x represents the original sample data, x min Indicates the minimum sample value; x max Indicates the maximum sample value; Represents the sample value after normalization.

4. The method for short-term multivariate load forecasting of an integrated energy system integrating a deep learning combined model as claimed in claim 1, characterized in that: The serial-parallel superimposed multi-scale convolutional neural network includes two convolutional neural networks.

5. The method for short-term multivariate load forecasting of an integrated energy system integrating a deep learning combined model as claimed in claim 1, characterized in that: A bidirectional gated recurrent unit neural network is used to learn the time series characteristics of load data and iteratively extract global information; the bidirectional gated recurrent unit neural network is provided with two layers of bidirectional gated recurrent units, the number of hidden layer neurons of each layer of bidirectional gated recurrent units is determined by an adaptive multi-strategy grey wolf optimization algorithm, and a Dropout layer is introduced into the hidden layer.

6. The method for short-term multivariate load forecasting of an integrated energy system integrating a deep learning combined model as claimed in claim 5, characterized in that: The calculation expression of the gated recurrent unit is as follows: z t =σ(W iz x t +b iz +W hz h t-1 +b ht ) r t =σ(W ir x t +b ir +W hr h t-1 +b hr ) h t =tanh[W in x t +b in +r t ⊙(W hn h t-1 +b hn )] h t =(1-z t )πh t-1 +with t ⊙h t Where: h t-1 is the hidden state at time t-1; x t is the input at time t; σ is the sigmoid activation function; r t is the reset gate; t is the update gate; h t is the final hidden state; W is the coefficient matrix; ⊙ is the Hadamard product; b is the bias coefficient; The bidirectional gated recurrent unit is composed of two-way gated recurrent units. The time series data is calculated in the two gated recurrent units in opposite directions, which are responsible for processing historical information and future information respectively. For an input sequence X = x1, x2, ..., x T , the forward computation of the bidirectional gated recurrent unit is expressed as: Where: and are the hidden states from left to right and from right to left, respectively, x t is the tth element of the input sequence; On this basis, the hidden states in the two directions are fused to obtain the final hidden state: Pass the hidden state to a fully connected layer to get the output y t : y t =softmax(Wh t +b) 7. The method for short-term multivariate load forecasting of an integrated energy system integrating a deep learning combined model as claimed in claim 6, characterized in that: The attention mechanism is a multi-head attention mechanism.

8. The method for short-term multivariate load forecasting of an integrated energy system integrating a deep learning combined model as claimed in claim 5, characterized in that: The adaptive multi-strategy gray wolf optimization algorithm is an improvement on the traditional gray wolf optimization algorithm. The improvements are as follows: (1) Initialization population of chaos model: The infinite folding SIN chaos model is used to initialize the population, and its expression is: (2) Nonlinear convergence factor: A nonlinear convergence factor is used to adapt and optimize the chaotic model. The expression of the nonlinear convergence factor is: Where: a init with a fin are the initial values ​​of the convergence factors respectively; (3) Dimensional learning individual update: The position update method based on dimensional learning between wolf individuals is introduced to improve the search ability of population individuals. The specific strategy is: First, the position X of each individual gray wolf i is calculated using the Euclidean distance i (t) and X i～G (t+1) domain range R i (t) is: R i (t)=||X i (t)-X i～G (t+1)|| Secondly, according to the R of each individual gray wolf i (t), construct the neighboring gray wolf group N of each individual gray wolf i (t) is: On this basis, multi-neighbor learning is performed, which is expressed as: X i～D (t+1)=X i (t)+rand·(X n (t)-X r (t)) Finally, the new position of the individual is updated according to the fitness value, which is expressed as: (4) Generalized reverse learning variation: A generalized reverse learning strategy is used to perturb the position of the leader wolf at the end of each search, which is expressed as: Where, X i a / b / d is the j-th dimension of the position of the leading wolves a, b and d, and k is the proportional factor; after the leading wolves a, b and d are disturbed, the fitness values ​​before and after the disturbance are compared to determine whether the leading wolves a, b and d update their positions.

9. The method for short-term multivariate load forecasting of an integrated energy system integrating a deep learning combined model as claimed in claim 1, characterized in that: The evaluation index is: root mean square error E RMSE , mean absolute error E MAE And the average relative error E MAPE , the specific calculation is as follows: Where: n represents the number of predicted loads, y i represents the actual value of the i-th load, represents the predicted value of the i-th load.

Images