Power Load Forecasting Method and System for Energy Storage System

Abstract

The present application relates to the field of energy storage technologies, and discloses a method and system for predicting the power load of an energy storage system. The method for predicting the power load of the energy storage system in the present application includes: obtaining the load sequence and the meteorological sequence for the current time period; determining whether there is a sequence in the historical load sequence of the energy storage system whose similarity to the load sequence exceeds a preset threshold; if not, obtaining a preliminary load prediction sequence for a future time period after the current time period according to the load sequence and the meteorological sequence; correcting the preliminary load prediction sequence to obtain a first load prediction sequence; if so, obtaining the first load sequence with the highest similarity, obtaining the actual load sequence and the actual meteorological sequence for the target time period after the first load sequence; obtaining the predicted meteorological sequence corresponding to the future time period; calculating the meteorological difference sequence between the actual meteorological sequence and the predicted meteorological sequence; obtaining the load difference sequence according to the meteorological difference sequence, and superimposing it on the actual load sequence to obtain a second load prediction sequence.

Description

Technical Field

[0001] This application relates to the technical field of energy storage, and particularly relates to a method and system for predicting the electrical load of an energy storage system. Background Art

[0002] Virtual Power Plant (VPP) and Microgrid are two important components in modern power systems. They play important roles in improving energy utilization efficiency, enhancing grid flexibility and reliability; the energy storage system can provide rapid response and regulation capabilities for the virtual power plant and microgrid, and provide stable electrical energy for the virtual power plant and microgrid. The load prediction of the virtual power plant and microgrid can optimize the energy dispatching of the energy storage system, and play an important role in improving the operation reliability and energy utilization efficiency of the energy storage system, and enhancing grid flexibility and reliability.

[0003] With the development of technology and the growth of demand, users have put forward higher requirements for the energy dispatching of the energy storage system. Therefore, higher prediction accuracy requirements are put forward for the load prediction technology in the virtual power plant or microgrid. However, the existing load prediction algorithms still have the problem of low load prediction accuracy, which affects the energy utilization efficiency of the energy storage system. Summary of the Invention

[0004] The purpose of the embodiments of this application is to provide a method and an energy storage system for predicting the electrical load of an energy storage system, so as to improve the accuracy of load prediction of the energy storage system.

[0005] To solve the above technical problems, an embodiment of the present application provides a method for predicting the power load of an energy storage system, including: obtaining the load sequence of the energy storage system in the current period, and obtaining the meteorological sequence of the current period; the meteorological sequence of the current period is a sequence composed of meteorological data corresponding to multiple moments, and the meteorological data at least includes: temperature, humidity, air pressure, weather type; wherein, the weather type is represented in the form of numerical coding; determining whether there is a load sequence in the historical load sequence of the energy storage system whose similarity to the load sequence exceeds a preset threshold; if not, obtaining a preliminary load prediction sequence for a future period after the current period according to the load sequence and the meteorological sequence; correcting the preliminary load prediction sequence to obtain a first corrected load prediction sequence; if so, obtaining a first load sequence with the highest similarity to the load sequence in the historical load sequence, obtaining the actual load sequence and the actual meteorological sequence of a target period after the period where the first load sequence is located; obtaining the predicted meteorological sequence corresponding to the future period after the current period; calculating the meteorological difference sequence between the actual meteorological sequence and the predicted meteorological sequence; obtaining a load difference sequence according to the meteorological difference sequence; and superimposing the load difference sequence on the actual load sequence to obtain a second load prediction sequence corresponding to the future period.

[0006] An embodiment of the present application further provides an energy storage system, including: an energy storage device, an energy storage bidirectional converter, and an energy management module. The energy management module is used to implement the method for predicting the power load of the energy storage system as described above, and the energy management module is further used to adjust the charge and discharge strategy of the energy storage system according to the first load prediction sequence or the second load prediction sequence.

[0007] In some embodiments, the obtaining a preliminary load prediction sequence for a future period after the current period according to the load sequence and the meteorological sequence includes: jointly using the load sequence and the meteorological sequence as the input sequence of a prediction model, so that the prediction model outputs the preliminary load prediction sequence for the future period; wherein, the prediction model includes an encoder and a decoder. The encoder includes multiple stacked encoder layers, a residual connection module and a layer normalization module arranged after the stacked encoder layers. Each encoder layer includes a multi-head self-attention mechanism module and a first feed-forward neural network module; the decoder includes a masked self-attention mechanism module, a frequency domain enhancement module, an attention mechanism module, and a second feed-forward neural network module connected in sequence; the frequency domain enhancement module performs moving average decomposition on the load sequence to obtain a low-frequency trend part and a high-frequency trend part, and randomly selects a fixed number of frequency components by applying discrete Fourier transform or wavelet transform as the base frequency for the attention mechanism module to calculate.

[0008] In some embodiments, the training process of the prediction model includes: obtaining an initial training sequence, where the initial training sequence includes feature data corresponding to the most recent N moments, and the feature data corresponding to each moment includes the load data at the corresponding moment and the meteorological data at the corresponding moment; N is an integer greater than 0; performing normalization processing on the initial training sequence to obtain a target training sequence, and the formula for the normalization processing is: ; where x is a feature data in the initial training sequence, min(X) is the minimum feature data in the initial training sequence, max(X) is the maximum feature data in the initial training sequence, and x' is a feature data in the normalized target training sequence; inputting the target training sequence into the prediction model for training to obtain the trained prediction model.

[0009] In some embodiments, using the load sequence and the meteorological sequence as the input sequence of the prediction model so that the prediction model outputs the preliminary load prediction sequence for the future period includes: the encoder receives the input sequence, and the length of the input sequence is M, where M is an integer greater than 0; the encoder extracts the features of the input sequence layer by layer to obtain the feature representation of the input sequence ; the formula of the encoder is: ; X is a feature data in the input sequence, and one feature data is composed of the load data and the meteorological data at the same moment, represents the mapping result of the encoder for the input feature data X; the decoder generates the preliminary load prediction sequence for the future period according to the feature representation Zenc and the feature data X at the prediction start point of the future period feat ; the formula of the decoder is: ; Y is the preliminary load prediction sequence, represents the mapping result of the decoder for the feature representation Zenc and the feature data X at the prediction start point of the future period feat of.

[0010] In some embodiments, the multi-head self-attention mechanism is provided with h independent attention heads in parallel, and each attention head has its own first projection matrix W i Q 、W i K 、W i V , where h is an integer greater than 1, and the first projection matrix W i Q 、W i K 、W i Vare respectively used for linearly transforming the query, key, and value; each of the attention heads maps the input sequence according to its respective first projection matrices \(W\) i Q , \(W\) i K , \(W\) i V to obtain the attention output \(Q\) i ; concatenate the \(h\) attention outputs \(Q\) i in the feature dimension to obtain the concatenated output \(Concat(O1, O2, \ldots, O\) h ); map the concatenated output \(Concat(O1, O2, \ldots, O\) O back to the original dimension through the second projection matrix \(W\) h ), and the calculation formula for the original dimension is: ; \(Q\) is the query vector, \(K\) is the key vector, and \(V\) is the value vector; the first feed-forward neural network module outputs the mapping \(F(x)\) corresponding to each feature data \(x\) in the initial training sequence according to the original dimension ; the residual connection module obtains the residual connection output according to the mapping \(F(x)\), and the calculation formula for the residual connection output is: ; the layer normalization module performs layer normalization on the residual connection output to obtain , and represents the result of performing layer normalization on ; the layer normalization module is also used to normalize the input of each encoder layer to zero mean and unit variance. In some embodiments, the masked self-attention mechanism module includes a mask matrix \(M\), where, for the query vector and key vector of the decoder, when \(j \leq i\), \(M\)

[0011] ij = 0, indicating that the position \(i\) corresponding to the key vector pays attention to the information at the position \(j\) corresponding to the query vector, and when \(j > i\), \(M\) ij ij = -\infty, indicating that the position \(i\) corresponding to the key vector cannot pay attention to the information at the position \(j\) corresponding to the query vector; the calculation formula for the masked self-attention mechanism module is: ; where, represents the output result of the masked self-attention mechanism module when the input query vector, key vector, and value vector are \(Q\), \(K\), and \(V\) in sequence, \(QK\) T represents the dot product of vector \(Q\) and vector \(K\), and \(d\) k is the dimension of vector \(K\), used to scale the result of \(QK\) T ), Indicates The value applied to the Softmax function.

[0012] In some embodiments, before correcting the preliminary load prediction sequence to obtain the corrected first load prediction sequence, it includes: obtaining the predicted meteorological sequence corresponding to the future time period; estimating the influence deviation of the predicted meteorological sequence on the preliminary load prediction sequence to obtain a meteorological correction value sequence; obtaining an error correction value sequence corresponding to the preliminary load prediction sequence according to the load sequence and the meteorological sequence; the correcting the preliminary load prediction sequence to obtain the corrected first load prediction sequence includes: obtaining the corrected first load prediction sequence according to the preliminary load prediction sequence, the meteorological correction value sequence, and the error correction value sequence.

[0013] In some embodiments, the estimating the influence deviation of the predicted meteorological sequence on the preliminary load prediction sequence to obtain a meteorological correction value sequence includes: using the predicted meteorological sequence as the input of a meteorological correction model, so that the meteorological correction model outputs the meteorological correction value sequence; wherein, the meteorological correction model includes a plurality of parallel convolutional kernels, and the meteorological correction model further includes a non-linear activation layer and a pooling layer; the width and dilation rate of each convolutional kernel are different; the way for the meteorological correction model to obtain the meteorological correction value sequence includes: each convolutional kernel sequentially extracts features from each meteorological data in the predicted meteorological sequence to obtain an initial meteorological feature f j ; The initial meteorological feature f j is obtained through the following expression: ; Indicates the convolution operation on when the width of the convolutional kernel is k and the dilation rate is d is the predicted meteorological sequence, W(t:t + H)=[w1, w2,..., w H , H is the number of prediction steps, and w1, w2,..., w H are the meteorological data at different times respectively; the meteorological correction model also fuses the initial meteorological features f j output by each convolutional kernel to obtain a comprehensive meteorological feature corresponding to each meteorological data ; The comprehensive meteorological feature is refined through the non-linear activation layer and the pooling layer to obtain the meteorological correction value sequence with a length of H, and each element in ΔYmet is the meteorological correction value corresponding to each meteorological data.

[0014] In some embodiments, obtaining the error correction value sequence corresponding to the preliminary load prediction sequence according to the load sequence and the meteorological sequence includes: using the load sequence and the meteorological sequence as the input sequences of an error learning model, so that the error learning model outputs the error correction value sequence; the loss function of the error learning model is: , where e(t) is the historical error correction value sequence learned by the error learning model, y(t) is the historical true load sequence, and is the historical predicted load sequence.

[0015] In some embodiments, the error learning model is a deep residual network model; the deep residual network model includes L fully connected layers, where L is an integer greater than 1; each fully connected layer learns and corrects the error part not captured by the previous fully connected layer. The output of the -th fully connected layer of the deep residual network model satisfies the following formula: , is greater than 0 and less than L; where, represents the input feature of the -th fully connected layer at time t + i. The input feature of the -th fully connected layer includes the output of the previous fully connected layer. The input feature of the first fully connected layer is the input sequence of the error learning model. is the parameter of the -th fully connected layer, is the correction term mapped based on and . After the L fully connected layers of the deep residual network model are stacked, the error correction value sequence output by the L-th fully connected layer is calculated by the formula: .

[0016] In some embodiments, obtaining the corrected first load prediction sequence according to the preliminary load prediction sequence, the meteorological correction value sequence, and the error correction value sequence includes: obtaining a meteorological correction prediction value sequence according to the preliminary load prediction sequence and the meteorological correction value sequence ; obtaining a reference prediction value sequence according to the preliminary load prediction sequence and the error correction value sequence . The calculation formula of the reference prediction value sequence is: , is the preliminary load prediction sequence; using the meteorological correction prediction value sequence, the reference prediction value sequence, and the auxiliary feature sequence as the inputs of the XGBoost gradient boosting tree model, so that the XGBoost gradient boosting tree model outputs the first load prediction sequence; the auxiliary feature sequence includes the key meteorological elements of the current time period, the key meteorological elements of the future time period, the actual load error sequence of the previous prediction cycle, and the time index feature of the future time period; the XGBoost gradient boosting tree model includes multiple regression trees; the XGBoost gradient boosting tree model aims to minimize the prediction error and the objective function of the XGBoost gradient boosting tree model is: ; where is the input of the loss function , is the result calculated when i is the true load value of the i-th sample, is the predicted load value of the i-th sample; is the regularization term, and f k represents the k-th regression tree, and T is the total number of the regression trees, The calculation formula of is: represents the predicted value of the k-th regression tree for the input feature vector x i of the i-th sample.

[0017] In some embodiments, the way for the XGBoost gradient boosting tree model to obtain the first load prediction sequence includes: the XGBoost gradient boosting tree model obtains the bias term, the first fusion weight of the reference prediction value sequence, and the second fusion weight of the meteorological correction prediction value sequence according to the auxiliary feature sequence; fusing the reference prediction value sequence and the meteorological correction prediction value sequence according to the first fusion weight , the second fusion weight , and the bias term to obtain the first load prediction sequence , and the calculation formula of the first load prediction sequence is: ; where represents the reference prediction value ， represents the meteorological correction prediction value sequence.

[0018] In some embodiments, the Gaussian similarity function is used to calculate the similarity of the load sequences in the historical load sequence, and the Gaussian similarity function is: ; where is a scale parameter set according to the historical fluctuation amplitude, is from the time point -M + 1 to the historical load sequence, is the load sequence from time point t - M + 1 to t, , is the load data corresponding to a moment in the load sequence.

[0019] In some embodiments, calculating the meteorological difference sequence between the actual meteorological sequence and the predicted meteorological sequence The calculation formula is: ; where represents the actual meteorological sequence, represents the predicted meteorological sequence; obtaining the load difference sequence according to the meteorological difference sequence includes: inputting the meteorological difference sequence into a meteorological correction model or an error learning model to estimate the load difference sequence caused by the meteorological difference .

[0020] The technical solutions provided by the embodiments of the present application have at least the following advantages:

[0021] In the case that there is no load sequence in the historical load sequence of the energy storage system whose similarity to the load sequence of the current period exceeds the preset threshold, the preliminary load prediction sequence for the future period after the current period is obtained according to the load sequence and the meteorological sequence, and the preliminary load prediction sequence is corrected to obtain the first corrected load prediction sequence; in the case that there is a load sequence in the historical load sequence of the energy storage system whose similarity to the load sequence of the current period exceeds the preset threshold, the first load sequence with the highest similarity to the load sequence of the current period is obtained in the historical load sequence, the actual load sequence and the actual meteorological sequence of a target period after the period where the first load sequence is located are obtained, and the predicted meteorological sequence corresponding to the future period after the current period is obtained. Calculate the meteorological difference sequence between the actual meteorological sequence and the predicted meteorological sequence, obtain the load difference sequence according to the meteorological difference sequence, and superimpose the load difference sequence on the actual load sequence to obtain the second load prediction sequence corresponding to the future period; thus, when the similarity between the load sequence of the current period and the historical load sequence is relatively high, the load prediction is quickly performed according to the historical load sequence of the energy storage system, and the meteorological data is incorporated into the load prediction, taking into account both the efficiency and accuracy of the load prediction; when the similarity between the load sequence of the current period and the historical load sequence is relatively low, the load prediction is performed by correcting the initial load prediction sequence and incorporating the meteorological data, improving the accuracy of the load prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] One or more embodiments are exemplarily illustrated by the pictures in the corresponding drawings. These exemplary illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are represented as similar elements. Unless otherwise stated, the drawings in the figures do not constitute a scale limitation.

[0023] Figure 1 It is a schematic flowchart of a method for predicting the power load of an energy storage system according to an embodiment of the present application;

[0024] Figure 2 It is a schematic flowchart of a method for predicting the power load of an energy storage system according to another embodiment of the present application;

[0025] Figure 3 It is a schematic flowchart of each sub-step of step 207 according to another embodiment of the present application;

[0026] Figure 4 It is a schematic flowchart of obtaining the first load prediction sequence by the XGBoost gradient boosting tree model according to another embodiment of the present application. Detailed implementation manners

[0027] As can be seen from the background art, the existing load prediction algorithms for energy storage systems still have the problem of low load prediction accuracy.

[0028] Through analysis and research, it is found that: most of the existing load prediction models rely on historical load data for prediction, but often ignore the dynamic changes of real-time meteorological data. Since meteorological factors (such as temperature, humidity, air pressure, etc.) have a significant impact on the load, traditional models cannot effectively capture the impact of these factors on load changes, resulting in low prediction accuracy; secondly, most of the existing load prediction models do not have the ability to flexibly respond to real-time changes. They still use the same model in different time periods and scenarios, and it is difficult to adapt to the changes in load patterns, resulting in low prediction accuracy. For example, in special situations such as holidays and extreme weather, the prediction errors of traditional models are relatively large; furthermore, the existing load prediction models are generally static. Once the prediction results are output each time, subsequent predictions cannot be updated. Therefore, in actual operation, if new meteorological changes or load fluctuations occur, the system cannot adjust the prediction results in time, resulting in a decrease in the later prediction accuracy.

[0029] When there is no load sequence in the historical load sequence of the energy storage system whose similarity to the load sequence at the current time period exceeds a preset threshold, a preliminary load prediction sequence for a future time period after the current time period is obtained based on the load sequence and the meteorological sequence, and the preliminary load prediction sequence is corrected to obtain a first corrected load prediction sequence; when there is a load sequence in the historical load sequence of the energy storage system whose similarity to the load sequence at the current time period exceeds a preset threshold, the first load sequence with the highest similarity to the load sequence at the current time period is obtained in the historical load sequence, the actual load sequence and the actual meteorological sequence of a target time period after the time period where the first load sequence is located are obtained, and the predicted meteorological sequence corresponding to the future time period after the current time period is obtained. The meteorological difference sequence between the actual meteorological sequence and the predicted meteorological sequence is calculated, the load difference sequence is obtained based on the meteorological difference sequence, and the load difference sequence is superimposed on the actual load sequence to obtain a second load prediction sequence corresponding to the future time period; thus, when the similarity between the load sequence at the current time period and the historical load sequence is relatively high, the load prediction is quickly performed based on the historical load sequence of the energy storage system, and the meteorological data is incorporated into the load prediction, taking into account both the efficiency and accuracy of the load prediction; when the similarity between the load sequence at the current time period and the historical load sequence is relatively low, the load prediction is performed by correcting the initial load prediction sequence and incorporating the meteorological data, improving the accuracy of the load prediction.

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the embodiments of the present application will be described in detail below with reference to the accompanying drawings. However, those of ordinary skill in the art can understand that in the embodiments of the present application, many technical details are proposed to help the reader better understand the present application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical solutions required to be protected by the present application can still be implemented. The following division of each embodiment is for convenience of description and should not constitute any limitation on the specific implementation manner of the present application. The various embodiments can be combined with each other and cross-referenced without conflict.

[0031] An embodiment of the present application relates to a method for predicting the electric load of an energy storage system, as Figure 1 shown in the flowchart of the method for predicting the electric load of the energy storage system in this embodiment. The method for predicting the electric load of the energy storage system in this embodiment includes the following steps:

[0032] Step 101, obtain the load sequence of the energy storage system at the current time period and obtain the meteorological sequence at the current time period.

[0033] Specifically, the load sequence of the current period is a sequence composed of load data corresponding to multiple moments, and the meteorological sequence of the current period is a sequence composed of meteorological data corresponding to multiple moments. The meteorological data at least includes: temperature, humidity, air pressure, weather type (including sunny, rainy, cloudy, overcast, etc.); among them, the weather type is represented in the form of numerical coding, and continuous meteorological elements such as temperature, humidity, and air pressure can be directly used as numerical features.

[0034] In this embodiment, a multi-modal feature fusion method is adopted to integrate meteorological data into the load prediction of energy storage devices to improve the accuracy of load prediction. Therefore, after obtaining the load sequence of the current period and the meteorological sequence of the current period, for each time step, the load data and meteorological data corresponding to the corresponding moment are concatenated, so that the subsequent model can receive the information of these two modalities at the same time.

[0035] Step 102, determine whether there is a load sequence in the historical load sequence of the energy storage system whose similarity to the load sequence exceeds a preset threshold.

[0036] In this embodiment, if there is no load sequence in the historical load sequence of the energy storage system whose similarity to the load sequence exceeds the preset threshold, go to step 103; if there is a load sequence in the historical load sequence of the energy storage system whose similarity to the load sequence exceeds the preset threshold, go to step 105.

[0037] In this embodiment, by determining whether there is a load sequence in the historical load sequence whose similarity to the load sequence of the current period exceeds the preset threshold, different processing is performed. When the similarity between the load sequence of the current period and the historical load sequence is relatively high, the load prediction is quickly performed according to the historical load sequence of the energy storage system, that is, the prediction is performed by the method of historical similarity matching. When the similarity between the load sequence of the current period and the historical load sequence is relatively low, the load prediction is performed by correcting the initial load prediction sequence, that is, the prediction is performed by the method of model inference correction, taking into account both the efficiency and accuracy of load prediction.

[0038] Specifically, this embodiment can adopt a rolling prediction mechanism. For example, the load prediction result is updated every 15 minutes to reflect the latest load and meteorological information in real time. Rolling prediction enables the prediction to be refreshed accordingly whenever new information becomes available. For example, if the actual load trend deviates from the previous period's prediction, a new similar search will select a more matching new segment, and the prediction will correct the trajectory. This continuous update mechanism greatly improves the sensitivity of the prediction to real-time changes, can promptly reflect sudden fluctuations in the load, and improves the accuracy of the prediction.

[0039] The rolling prediction mechanism of this embodiment is implemented by the method of "historical similarity matching + model inference correction", that is, at each update moment, first find the load segment similar to the current scenario in history as a reference, that is, use the method of historical similarity matching for prediction. If there is no load segment similar to the current scenario in history, then combine the latest model correction to form the final prediction, that is, use the method of model inference correction for prediction.

[0040] In some embodiments, the similarity of the load sequences in the historical load sequence is calculated using a Gaussian similarity function. The Gaussian similarity function is:

[0041] ; where is a scale parameter set according to the historical fluctuation amplitude, is the historical load sequence from time point -M + 1 to , is the load sequence from time point t - M + 1 to t, , as the matching feature, is the load data corresponding to a moment in the load sequence.

[0042] Specifically, in the rolling prediction mechanism, for each rolling moment , first search the historical database for segments similar to the current load state. The current load state can be represented by a feature matching vector, and generally, the load sequence for a period of time before the current moment can be taken. For example, the load sequence of the most recent M 15-minute load values can be used to form the load sequence of the current moment, that is, the vector . Among them, when M is large, downsampling or key points (such as the maximum and minimum loads in the recent few hours) can be added to compress the features; by calculating the Gaussian similarity function between the vector and multiple historical load sequences respectively, obtain the similarity between the load sequence and each historical load sequence . The closer the similarity is to 1, the more similar the sequences are, and the closer it is to 0, the greater the difference. Then, determine whether there is a load sequence in the multiple historical load sequences whose similarity to the load sequence exceeds the preset threshold, so as to perform different load prediction processes respectively, taking into account both the efficiency and accuracy of load prediction.

[0043] Step 103, obtain the preliminary load prediction sequence for the future period after the current period according to the load sequence and the meteorological sequence.

[0044] Specifically, after obtaining the load sequence and the meteorological sequence in this embodiment, the load sequence and the meteorological sequence are fused as multi-modal features, and a preliminary load prediction sequence for a future period is obtained according to the multi-modal features fused from the load sequence and the meteorological sequence, so as to integrate meteorological data into the load prediction of the energy storage device and improve the accuracy of load prediction.

[0045] In some embodiments, step 103, that is, obtaining a preliminary load prediction sequence for a future period after the current period according to the load sequence and the meteorological sequence, can be implemented in the following manner: The load sequence and the meteorological sequence are jointly used as the input sequence of the prediction model, so that the prediction model outputs the preliminary load prediction sequence for the future period. Specifically, the multi-modal features fused from the load sequence and the meteorological sequence are input into the prediction model, and the prediction model can output the corresponding preliminary load prediction sequence for the future period according to the multi-modal features.

[0046] In some embodiments, the load sequence and the meteorological sequence are jointly used as the input sequence of the prediction model, so that the prediction model outputs the preliminary load prediction sequence for the future period, which can be implemented in the following manner: The encoder receives the input sequence, that is, the multi-modal features fused from the load sequence and the meteorological sequence. The length of the input sequence is M, and M is an integer greater than 0; the encoder extracts the features of the input sequence layer by layer to obtain the feature representation of the input sequence ; The formula of the encoder is: ; X is a feature data in the input sequence, and a feature data is composed of the load data and meteorological data at one moment, represents the mapping result of the encoder for the input feature data X; the decoder generates a preliminary load prediction sequence for the future period according to the feature representation Zenc and the feature data at the prediction starting point of the future period; the formula of the decoder is: ; where, is the feature data at the prediction starting point, and this feature data includes the load data and meteorological data at the prediction starting point. Y is the preliminary load prediction sequence, , Y represents the preliminary load prediction sequence starting from time t with a step size of H, and this preliminary load prediction sequence is used as the benchmark prediction for the subsequent correction process, represents the mapping result of the decoder for the feature representation Zenc and the feature data X feat at the prediction starting point of the future period.

[0047] The prediction model of this embodiment is an improvement on the FEDformer (Frequency Enhanced Decomposed Transformer) model structure. Its structure is based on the Transformer architecture and combines improvements such as frequency domain enhancement and residual network.

[0048] The improved FEDformer model in this embodiment adopts an encoder-decoder structure similar to the standard Transformer, including an encoder and a decoder. The encoder includes multiple stacked encoder layers, a residual connection module and a layer normalization module arranged after the stacked encoder layers. Each encoder layer includes a multi-head self-attention mechanism module and a first feed-forward neural network module; the decoder includes a masked self-attention mechanism module, a frequency domain enhancement module, an attention mechanism module, and a second feed-forward neural network module connected in sequence. The encoder is used to receive an input sequence of length , and the input sequence includes load data and meteorological data, which is represented as an input matrix after linear embedding, is the feature dimension, and the representation of the sequence is extracted layer by layer ; the decoder takes the encoder output and the known information at the prediction starting point as inputs to generate a prediction sequence steps (for example, 16 15-minute steps, that is, 4 hours) in the future . That is, the improved FEDformer model architecture captures the historical patterns at the current moment through the encoder and extrapolates the patterns to the future through the decoder.

[0049] The core of the traditional Transformer model lies in self-attention, which enables the model to efficiently focus on the correlations between different time steps in the sequence. For example, given the input ( is the sequence length, in the encoder ), it is first linearly mapped to a query , a key and a value through a trainable matrix, where , is the attention space dimension, and then the dot-product attention is calculated. The formula is as follows:

[0050] ; where, obtains the correlation score matrix between each time step, performs Softmax normalization on it row by row to obtain the attention weights, and then sums the weighted value vectors to obtain the output. Taking the attention calculation of the layer as an example, let , then the attention output , and its size is the same as the input . Through self-attention, the traditional Transformer model can automatically allocate attention weights according to the correlations. For example, the current load point may have a higher attention to the load at the same moment one day ago, so as to capture the daily cycle pattern.

[0051] Since a single attention head may not be sufficient to represent complex patterns, this embodiment adopts multi-head attention (Multi-Head Attention), that is, each encoder layer includes a multi-head self-attention mechanism module and a first feedforward neural network module, and the multi-head self-attention mechanism module of the encoder layer is set in parallel. independent attention heads, each with its own first projection matrix , h is an integer greater than 1, W i Q 、W i K 、W i V They are used to perform linear transformations on queries, keys, and values respectively.

[0052] Specifically, each attention head has its own learnable first projection matrix W i Q 、W i K 、W i V , used to perform linear transformations on queries, keys, and values, respectively. The first projection matrix W i Q 、W i K 、W i V The input query, key, and value vectors are mapped to different feature spaces. This way, the input sequence is projected into different subspaces, enabling each attention head to learn different feature representations. Each attention head can focus on different patterns or timescales in the input sequence, helping the forecasting model simultaneously capture multiple characteristics such as seasonal trends, diurnal cycles, and random fluctuations in load.

[0053] Specifically, each attention head is calculated based on its first projection matrix W i Q 、W i K 、W i V The attention output Q obtained by mapping the input sequence i , output h attentions to Q i Concatenate on the feature dimension to get the concatenated output Concat(O1, O2, ..., O h ), through the second projection matrix W O Concatenate the output Concat(O1,O2,…,O h ) is mapped back to the original dimension, the original dimension The calculation formula is: ; Q is the query vector, K is the key vector, and V is the value vector.

[0054] Specifically, each encoder layer further includes a first feed-forward neural network module connected to the outputs of multiple attention heads. The first feed-forward neural network module outputs the mapping F(x) corresponding to each feature data x in the original dimension . After that, to alleviate the vanishing gradient of the deep network and improve the feature transfer efficiency, the improved FEDformer model sets a residual connection module and a layer normalization module after each sub-layer, namely the multi-head attention mechanism and the first feed-forward neural network module. The residual connection module obtains the residual connection output according to the mapping , and the calculation formula of the residual connection output is: ; the layer normalization module performs layer normalization on the residual connection output to obtain , represents the result of performing layer normalization on ; the layer normalization module is used to normalize the input of each encoder layer to zero mean and unit variance. Through this recursive form of residual accumulation in this embodiment, each layer only needs to learn the incremental change, that is, the mapping , so as to realize the gradual refinement of the input features, and can also enable the model to better learn the layer-by-layer correction of the prediction error. The layer normalization module can ensure numerical stability after stacking the residuals, thus accelerating convergence.

[0055] The decoder in this embodiment includes a masked self-attention mechanism module, a frequency domain enhancement module, an attention mechanism module, and a second feed-forward neural network module connected in sequence. In the design process of the decoder, to avoid seeing future information and prevent future data from affecting the decoder, a mask mechanism is adopted for self-attention. The addition of the mask ensures that the prediction is strictly based on past data.

[0056] Specifically, the masked self-attention mechanism module includes a mask matrix M. Among them, for the query vector and the key vector of the decoder, when j ≤ i, M ij = 0, indicating that the position i corresponding to the key vector pays attention to the information of the position j corresponding to the query vector. When j > i, M ij = -∞, indicating that the position i corresponding to the key vector cannot pay attention to the information of the position j corresponding to the query vector; the calculation formula of the masked self-attention mechanism module is:

[0057] ; where Denotes the output result of the masked self-attention mechanism module when the input query vector, key vector, and value vector are Q, K, and V in sequence, QK T Denotes the dot product of the query vector Q and the key vector K, d k Is the dimension of the key vector, used to scale the result of QK T Of the result, Denotes taking As the input, the value of the Softmax function.

[0058] Specifically, for the query Query and key Key of the decoder's input sequence, when the position corresponding to the key Key is after the query Query position, its attention score is forced to be Or set to a very small negative value, so that the weight after Softmax is approximately 0. This ensures that when the decoder generates the th future load prediction, it will only refer to the information before the moment (including the historically known part and the first few predictions that have been generated), and will not misuse the future values after the moment Therefore, the masked attention mechanism module ensures the causality of the decoding process.

[0059] Specifically, the frequency domain enhancement module performs a moving average decomposition on the load sequence to obtain a low-frequency trend part and a high-frequency trend part, and applies the discrete Fourier transform or wavelet transform to randomly select a fixed number of frequency components as the base frequencies for the calculation of the attention mechanism module.

[0060] The frequency domain enhancement module is a unique improvement of FEDformer in this embodiment. It integrates frequency domain information in the attention calculation. By performing a moving average decomposition on the sequence, it obtains the trend and seasonal components, decomposing the original load sequence into a low-frequency trend part (representing the slow change or long-term trend of the load) and a relatively high-frequency seasonal / cyclic part (representing cyclic fluctuations such as daily cycles). In the frequency domain, the trend corresponds to components close to zero frequency (extremely low frequency band, representing smooth changes), while seasonal components such as daily cycles appear as significant components at specific non-zero frequencies (for example, a 24-hour cycle will form a spectral peak at the corresponding cycle frequency), that is, the low-frequency band in the spectrum reflects the trend of the load, and the specific medium and high-frequency bands reflect the main periodic (seasonal) fluctuations. After obtaining the low-frequency trend part and the high-frequency trend part, the frequency domain enhancement module applies the discrete Fourier transform (DFT) or wavelet transform to extract the main frequency components. In the FEDformer model, by performing the above decomposition and spectral analysis on the load sequence, only several important frequency bases are selected for calculation, thereby reducing the impact of the time series length on the calculation.

[0061] The frequency-domain enhancement module of this embodiment can capture the periodic structure in the load sequence, improving the efficiency and stability of long-sequence prediction; by embedding periodic features in the attention mechanism, the model becomes more sensitive to periodic changes during long-term prediction, thereby improving the performance of traditional Transformer in long-sequence prediction.

[0062] In some embodiments, the training process of the prediction model includes: obtaining an initial training sequence, where the initial training sequence includes the feature data corresponding to the most recent N moments, and the feature data corresponding to each moment includes the load data and the meteorological data at the corresponding moment; N is an integer greater than 0; performing normalization processing on the initial training sequence to obtain a target training sequence, and the formula for normalization processing is: ; where x is a feature data in the initial training sequence, min(X) is the minimum feature data of the initial training sequence, max(X) is the maximum feature data of the initial training sequence, and x' is a feature data of the normalized target training sequence; inputting the target training sequence into the prediction model for training to obtain the trained prediction model.

[0063] The normalization processing of this embodiment uniformly normalizes the original features from different sources and with different dimensions (such as Min-Max normalization or Z-score normalization), mapping each feature value to a similar numerical range.

[0064] Through normalization, this embodiment eliminates the dimension difference, accelerates the model convergence, and avoids certain features from dominating the model training due to too large numerical ranges. Especially for data with different orders of magnitude such as load values and meteorological values, normalization can ensure the gradient balance of each input dimension of the FEDformer model.

[0065] Specifically, in the training process of the prediction model of this embodiment, the initial training sequence is a multi-modal feature that combines load data and meteorological data. By introducing multi-modal features in the model training process, the model can learn cross-modal associations such as "high temperature leads to an increase in air-conditioning load", improving the accuracy of model prediction. In some embodiments, the multi-modal features input to the prediction model can also include time index features, such as hours, day of the week, holidays, etc., which are provided to the prediction model through periodic embedding or encoding methods, enabling the prediction model to better depict the daily and weekly cycle patterns of the load. The prediction model of this embodiment can autonomously learn the association relationships between different features and the load from multi-modal inputs, improving the accuracy and robustness of prediction.

[0066] Specifically, when training the above prediction model in this embodiment, it is necessary to determine a time window size, that is, to determine the length of the historical sequence that the model needs to learn. In this embodiment, an appropriately long window can be selected through experiments to cover the main periodic features and related influences. For example, load data and meteorological data from the past 1 day or longer (such as the same quarter within a week) are used as inputs, so that the prediction model can capture the daily and nightly load change patterns and recent trends. At the same time, in order to predict the demand updated every 15 minutes in a rolling manner, the time window cannot be too long to avoid increasing the calculation delay.

[0067] To achieve the accuracy and real-time performance of the prediction model, assume that the data of the most recent time instants are selected as the input feature sequence of the model , where each time instant includes load data and various meteorological data. Through the method of global search and cross-validation, the optimal value can be determined to make the window length cover the main patterns and avoid redundant noise. At the same time, before model training, that is, in the preprocessing, resampling and alignment are performed on the data within the time window to ensure that the historical load sequence and the historical meteorological sequence are precisely corresponding in time, and the sliding window is used to smooth the missing values and outliers to improve the data quality.

[0068] Step 104: Modify the preliminary load prediction sequence to obtain the first modified load prediction sequence.

[0069] After obtaining the preliminary load prediction sequence in this embodiment, the preliminary load prediction sequence is modified, thereby improving the accuracy of the obtained first load prediction sequence and the accuracy of load prediction.

[0070] Step 105: Obtain the first load sequence with the highest similarity to the load sequence in the historical load sequence, and obtain the actual load sequence and the actual meteorological sequence of a target period after the period where the first load sequence is located.

[0071] After determining in step 102 above that there is a load sequence in the historical load sequence of the energy storage system whose similarity to the load sequence exceeds the preset threshold, this embodiment proceeds to step 105, obtains the first load sequence with the highest similarity to the load sequence in the historical load sequence, and obtains the actual load sequence and the actual meteorological sequence of a target period after the period where the first load sequence is located.

[0072] This embodiment calculates the similarity of multiple segments in the historical database through the Gaussian similarity function algorithm, selects several with the highest similarity as candidates, and selects one historical load sequence with the highest similarity called the similar day sequence, and the true load of the future H steps, such as 4 hours, after the time instant of this similar day sequence Take it out as the inference load sequence at the current moment For the preliminary estimation, that is, to obtain the actual load sequence of the target period after the period where the first load sequence is located. This embodiment draws on the idea of "similar days", that is, if the current load curve shape has appeared in history, then the subsequent trend is likely to be similar, providing a prediction benchmark based on historical prior for the inference load sequence.

[0073] Specifically, only considering the similarity between the historical meteorological sequence and the load sequence at the current moment can only obtain a rough load result, and the accuracy of load prediction still cannot be guaranteed. It is also necessary to consider the differences between the current and historical external conditions, such as the impact of meteorological data on the load. Therefore, in this embodiment, while obtaining the actual load sequence of a target period after the period where the first load sequence is located, the actual meteorological sequence corresponding to the target period is also obtained.

[0074] Step 106: Obtain the predicted meteorological sequence corresponding to the future period after the current period; calculate the meteorological difference sequence between the actual meteorological sequence and the predicted meteorological sequence.

[0075] Specifically, this embodiment considers the impact of meteorological data on the load, and also obtains the predicted meteorological sequence corresponding to the future period after the current period. After obtaining the actual meteorological sequence corresponding to the above-mentioned target period and the predicted meteorological sequence corresponding to the future period, the meteorological difference sequence between the actual meteorological sequence and the predicted meteorological sequence is calculated, so as to subsequently obtain the load difference sequence according to the meteorological difference sequence, thereby taking into account the impact of meteorological data on the load in the historical similarity matching method and improving the accuracy of load prediction in the historical similarity matching process.

[0076] In some embodiments, step 107 is to calculate the meteorological difference sequence between the actual meteorological sequence and the predicted meteorological sequence The calculation formula is:[[]] ; where represents the actual meteorological sequence, represents the predicted meteorological sequence; that is, this embodiment calculates the difference between each meteorological data in the actual meteorological sequence and each meteorological data in the predicted meteorological sequence, and these differences constitute the meteorological difference sequence .

[0077] Step 107: Obtain the load difference sequence according to the meteorological difference sequence.

[0078] Specifically, after obtaining the above-mentioned meteorological difference sequence in this embodiment, the load difference sequence is obtained according to the meteorological difference sequence, so as to calculate the impact of meteorological data on the load, that is, the load difference sequence, thereby improving the accuracy of the obtained load prediction sequence.

[0079] In some embodiments, step 108, i.e., obtaining the load difference sequence according to the meteorological difference sequence, is implemented in the following manner: The meteorological difference sequence is input into the meteorological correction model or the error learning model to estimate the load difference sequence caused by the meteorological difference . Among them, the meteorological correction model and the prediction model are the models mentioned in the subsequent model inference and correction method, and the meteorological correction model and the prediction model will be described in detail later. In this embodiment, the meteorological correction model and the prediction model used in the subsequent model inference and correction method are applied to the historical similarity matching method, so as to improve the utilization rate of the models mentioned in the subsequent model inference and correction method, reduce the design cost, and also improve the accuracy of the load prediction result.

[0080] Step 108: The load difference sequence is superimposed on the actual load sequence to obtain the second load prediction sequence corresponding to the future time period.

[0081] Specifically, in this embodiment, after obtaining the load difference sequence, the load difference sequence is superimposed on the actual load sequence to obtain the second load prediction sequence corresponding to the future time period, so as to consider the influence of meteorology on the load during the load prediction process and improve the accuracy of the load prediction.

[0082] In this embodiment, by obtaining the meteorological records of the historical load sequence with the highest similarity, the actual meteorological sequence of a target time period after obtaining the historical similarity sequence is compared with the predicted meteorological sequence of the future time period after the current time period. For example, represents the actual meteorological sequence of a target time period after the historical load sequence with the highest similarity, represents the predicted meteorological sequence of the future time period after the current time period, and the difference between the two is calculated . For example, if the actual meteorological sequence is cloudy and the predicted meteorological sequence of the future time period after the current time period forecasts sunny, then the difference is positive in terms of solar radiation. After that, the meteorological difference sequence is input into the meteorological correction model or the error learning model to estimate the load difference sequence caused by the meteorological difference . In this embodiment, the aforementioned trained meteorological correction model and error learning model can be used to execute the process of obtaining the load difference sequence according to the meteorological difference sequence. Since the meteorological correction model and the prediction model have learned the relationship between meteorology and load error, they can be generalized to new scenarios, thereby reducing the design cost and also improving the accuracy of the load prediction result.

[0083] Specifically, the error learning model can be set as , which means that the load difference sequence is output according to the input meteorological difference sequence. Then, the formula for the error learning model to calculate the load difference sequence is:

[0084] ; Among them, represents a load difference sequence starting from time t with a step size of H, represents a meteorological difference sequence starting from time t with a step size of H.

[0085] After that, the formula for obtaining the second load prediction sequence corresponding to the future period by superimposing the load difference sequence on the actual load sequence is: ; Among them, represents the actual load sequence starting from time t with a step size of H, represents the second load prediction sequence starting from time t with a step size of H.

[0086] In this embodiment, after obtaining the similar highest historical load sequence, if the current predicted meteorological conditions are different from the meteorological conditions corresponding to the similar history, the model will up - regulate or down - regulate the original historical load pattern. For example, if it is hotter currently than in history, will be positive as a whole, making the prediction curve higher than the pure historical curve; if it is colder currently than in history, will be negative as a whole, making the prediction curve lower than the pure historical curve.

[0087] The load prediction method of the energy storage system in this embodiment can quickly perform load prediction according to the historical load sequence of the energy storage system when the load sequence in the current period has a high similarity with the historical load sequence, and incorporates meteorological data for load prediction, taking into account both the efficiency and accuracy of load prediction; when the load sequence in the current period has a low similarity with the historical load sequence, it performs load prediction by modifying the initial load prediction sequence and incorporating meteorological data, improving the accuracy of load prediction. That is to say, this embodiment combines the meteorological data in the open - source meteorological large model with traditional load prediction, thereby improving the prediction accuracy, especially being able to more accurately predict load fluctuations in the face of meteorological changes.

[0088] Another embodiment of this application relates to a method for predicting the electrical load of an energy storage system. As Figure 2 shown, it is a schematic flowchart of the method for predicting the electrical load of the energy storage system in this embodiment. The method for predicting the electrical load of the energy storage system in this embodiment includes the following steps:

[0089] Step 201, obtain the load sequence of the energy storage system in the current period, and obtain the meteorological sequence in the current period.

[0090] Step 202, determine whether there is a load sequence in the historical load sequence of the energy storage system whose similarity to the load sequence exceeds a preset threshold.

[0091] In this embodiment, since there is no load sequence in the historical load sequence of the energy storage system whose similarity to the load sequence exceeds the preset threshold, step 203 is entered. In the case where there is a load sequence in the historical load sequence of the energy storage system whose similarity to the load sequence exceeds the preset threshold, step 208 is entered.

[0092] Step 203: Obtain a preliminary load prediction sequence for a future period after the current period according to the load sequence and the meteorological sequence.

[0093] Step 204: Obtain a predicted meteorological sequence corresponding to the predicted future period.

[0094] Step 205: Estimate the influence deviation of the predicted meteorological sequence on the preliminary load prediction sequence to obtain a meteorological correction value sequence.

[0095] In this embodiment, by obtaining a predicted meteorological sequence corresponding to a future period after the current period, calculating the influence deviation of the predicted meteorological sequence on the preliminary load prediction sequence, and obtaining a meteorological correction value sequence, the influence of the meteorological sequence is introduced during the correction process of the preliminary load prediction sequence, thereby improving the accuracy of the final load prediction result.

[0096] In some embodiments, estimating the influence deviation of the predicted meteorological sequence on the preliminary load prediction sequence to obtain a meteorological correction value sequence includes: using the predicted meteorological sequence as the input of a meteorological correction model so that the meteorological correction model outputs a meteorological correction value sequence; wherein, the meteorological correction model includes a plurality of parallel convolutional kernels, and the meteorological correction model further includes a non-linear activation layer and a pooling layer; the width and dilation rate of each convolutional kernel are different; the way for the meteorological correction model to obtain the meteorological correction value sequence includes: each convolutional kernel sequentially extracts features from each meteorological data in the predicted meteorological sequence to obtain an initial meteorological feature f j ; the initial meteorological feature f j is obtained through the following expression: ; represents the convolution operation of when the convolutional kernel width is k and the dilation rate is d, is the predicted meteorological sequence, W(t:t+H)=[w1,w2,...,w H , H is the prediction step, and w1,w2,...,w H are the meteorological data at different times respectively; the initial meteorological features f j output by each convolutional kernel are fused to obtain a comprehensive meteorological feature corresponding to each meteorological data; the comprehensive meteorological feature is refined through a non-linear activation layer and a pooling layer to obtain a meteorological correction value sequence with a length of H, and each element in ΔYmet is the meteorological correction value corresponding to each meteorological data.

[0097] Since the influence of meteorological factors on load is complex and diverse, for example, a sudden rise in temperature may cause a sharp increase in the load of electrical appliances such as air conditioners, and the change of weather from sunny to rainy may reduce the light, resulting in an increase in the load of electrical appliances such as lighting. Therefore, in this embodiment, a meteorological correction module is added. The meteorological correction model of this embodiment can be modeled based on MSCNN (Multi-Scale Convolutional Neural Network), and the multi-scale convolutional neural network is used to extract features and estimate the influence of meteorological data. For example, when obtaining the latest meteorological forecast every 4 hours, the load prediction is corrected.

[0098] Specifically, the MSCNN of this embodiment can extract the change patterns of meteorological elements at different time scales by introducing convolutional kernels and dilation rates of different sizes. For the input meteorological sequence (for example, the 15-minute-by-15-minute meteorological forecast sequence for the next 4 hours, multi-dimensional including temperature, humidity, etc.) in this embodiment, multiple parallel convolutional branches are constructed. Among them, some convolutional branches use small convolutional kernels (such as length 3-5) to capture short-term drastic changes, and some other convolutional branches use larger convolutional kernels or dilated convolutions (Dilated CNN) to obtain the trend influence of longer periods.

[0099] Specifically, each convolutional branch has its own convolutional receptive field, which can focus on weather changes at different scales. For example, the small receptive field convolution can capture the instantaneous influence of sudden weather mutations (such as thunderstorms) on the load, and the large receptive field convolution extracts the slow effect of the daily temperature change trend on the load.

[0100] Specifically, in this embodiment, after obtaining the initial meteorological feature f of each convolutional output j the initial meteorological features extracted by each convolutional branch are concatenated or weighted and fused in the channel dimension to form a comprehensive meteorological feature representation . The fusion method can be to directly concatenate and then connect a fully connected layer or perform dimensionality reduction through convolution to enable the meteorological correction model to automatically learn the importance of features at each scale; or weights manually set for different scales of features can be used to achieve weighted averaging.

[0101] This embodiment can obtain the comprehensive meteorological feature by using a simple concatenation method , such as: , and then through a non-linear activation layer (such as ReLU) and a pooling layer, extract a meteorological influence sequence with a length of ; where Each element in it corresponds to a meteorologically induced load correction value at a future moment. The meteorological correction model of this implementation ensures that the model considers both rapid weather fluctuations and slow trends through multi-scale fusion. The output corresponding to the meteorological correction model can simultaneously reflect short-term anomalies and long-term deviations, thereby improving the accuracy of correcting the preliminary load sequence and enhancing the accuracy of load forecasting.

[0102] The meteorological correction value sequence output by the MSCNN model of this embodiment can be regarded as a factor or deviation amount for the preliminary load sequence, used to correct the preliminary load sequence. For example, if a load data in the preliminary load sequence is denoted as , output by the prediction model, and a meteorological data in the meteorological correction value sequence output by the meteorological correction module is denoted as , then a load prediction data in the final load prediction sequence that incorporates the meteorological impact is:

[0103] ; where can be regarded as the additive impact of the meteorological correction factor on the load data at this moment.

[0104] This embodiment can also consider multiplicative impacts and can change the factor effect to the form of . Since the relationship between load data and meteorological data is often approximately linear, the above-mentioned additive correction can more effectively correct and improve the accuracy of prediction.

[0105] During the training process of MSCNN in this embodiment, the deviation between historical meteorological data and historical load data can be used as samples to supervise the learning, enabling it to approximate the true "meteorologically induced load error". For example, when the actual occurrence of high temperature causes the load to be higher than the benchmark prediction, MSCNN will adjust the convolutional kernel weights to output a positive ; conversely, it outputs a negative deviation at low temperature, thereby making a targeted correction to the benchmark prediction during inference.

[0106] In practical applications, whenever new meteorological forecasts (for example, updated every 4 hours) are obtained, the latest future meteorological sequence is input into MSCNN to calculate the meteorological correction value sequence for the next few hours. Then, the meteorological correction value sequence is added to the preliminary load prediction sequence output by the prediction model within the current rolling cycle to obtain the first load prediction after real-time correction. It should be noted that if the previous meteorological forecasts have been considered in the previous prediction, then the meteorological correction is actually calculating the impact of the difference between the new and old meteorological forecasts on the load. When the change in the meteorological forecast is very small, will approach zero, indicating that no major adjustment is required; when significant changes occur in the forecast (such as a large increase in the expected temperature), a positive deviation is given to raise the load forecast. In this embodiment, the meteorological correction model ensures that the load forecast follows the update of meteorological information in real time, improving the accuracy of the load forecast.

[0107] Step 206, obtain the error correction value sequence corresponding to the preliminary load forecast sequence according to the load sequence and the meteorological sequence.

[0108] In this embodiment, by obtaining the load sequence and the meteorological sequence of the current period, and obtaining the error correction value sequence corresponding to the preliminary load forecast sequence according to the load sequence and the meteorological sequence, the influence of the meteorological sequence is introduced in the process of correcting the preliminary load forecast sequence, improving the accuracy of the final load forecast result.

[0109] In some embodiments, obtaining the error correction value sequence corresponding to the preliminary load forecast sequence according to the load sequence and the meteorological sequence includes: using the load sequence and the meteorological sequence as the input sequence of the error learning model, so that the error learning model outputs the error correction value sequence; the loss function of the error learning model is:

[0110] , where e(t) is the historical error correction value sequence learned by the error learning model, y(t) is the historical true load sequence, is the historical predicted load sequence.

[0111] In some embodiments, the error learning model is a deep residual network model (ResNet); the ResNet deep residual network model includes L fully connected layers, where L is an integer greater than 1; each fully connected layer learns and corrects the error part not captured by the previous fully connected layer. The output of the th fully connected layer of the ResNet deep residual network model satisfies the following formula:

[0112] , is greater than 0 and less than L; where, represents the input feature of the th fully connected layer. The input feature of the th fully connected layer includes the output of the previous fully connected layer. The input feature of the first fully connected layer is the input sequence of the error learning model. is the th fully connected layer parameter, is based on and mapped correction term; after stacking the L fully connected layers of the ResNet deep residual network model, the error correction value sequence The calculation formula is as follows: .

[0113] In this embodiment, in order to further improve the accuracy of load forecasting, an error learning module is introduced. The deep residual network model is used to model the prediction error, and the prediction deviation of the prediction model FEDformer is recursively learned and corrected. The core idea of the error learning model in this embodiment is to let the model learn the difference between the predicted value and the true value, and use this difference as the compensation for the preliminary load forecasting sequence, so as to improve the accuracy of the final prediction.

[0114] The goal of the error learning module is to predict the future error based on historical information, that is, the load sequence and the meteorological sequence ， Then use it to correct a load data in the preliminary load sequence , and the finally obtained load forecasting result is a load data in the first load forecasting sequence which is: .

[0115] The ResNet deep residual network model has the property of residual connection, and avoids the degradation of deep networks by learning residuals layer by layer. In the error learning model, each layer tries to learn and correct the error part that the previous layer fails to capture. For example, the first layer learns the deviation that the overall load is too high or too low, and the second layer learns more detailed periodic errors, and so on. Referring to the output of the th layer of the above-mentioned ResNet deep residual network model, which satisfies the formula. This formula reflects the idea of recursive residuals, that is, each layer adds a new correction term on the basis of the previously accumulated error estimate . At the same time, the ResNet deep residual network model ensures that even if the number of fully connected layers is large, the gradient can be effectively transmitted through the identity mapping, and the network can continuously accumulate and learn errors without losing stability.

[0116] The basic unit of the error learning model ResNet in this embodiment includes multiple fully connected layers, plus non-linear activation, and then the input is directly added to the output. For example, if two fully connected layers are used as the residual unit, for the input feature , the residual unit calculates , and then outputs as the input of the next layer. During the training process, we use the error calculated from historical data as the true value, and use the mean square error (MSE) loss to train ResNet to approximate these errors. The loss function can be:

[0117] , where is the number of training samples. In this embodiment, by minimizing , enabling the error learning model ResNet to learn to output an estimated to compensate for the model bias. It is worth mentioning that since the error is usually smoother and smaller in magnitude than the load data in the original load sequence, ResNet can more easily learn the pattern of the error. For example, if the model always underestimates the load in the early morning, ResNet can specifically learn this pattern and output a positive bias correction value in the early morning to offset the systematic error of the model.

[0118] In this embodiment, when making an actual prediction, information such as the load sequence and meteorological sequence before the current moment is input into the error learning model ResNet to obtain the sequence of error correction values for the next steps, i.e., . Then it is added to the preliminary load prediction sequence output by the prediction model FEDformer, i.e., , to obtain the corrected first load prediction sequence

[0119] In this way, the error learning model ResNet realizes the bias correction of the FEDformer prediction, improving the accuracy and robustness of the model.

[0120] Step 207: Obtain the corrected first load prediction sequence according to the preliminary load prediction sequence, meteorological correction value sequence, and error correction value sequence.

[0121] As Figure 3 shown, it is a schematic flowchart of each sub-step of step 207 of this embodiment. Step 207 is to obtain the corrected first load prediction sequence according to the preliminary load prediction sequence, meteorological correction value sequence, and error correction value sequence, including the following steps:

[0122] Step 2071: Obtain the meteorological correction prediction value sequence according to the preliminary load prediction sequence and meteorological correction value sequence.

[0123] Step 2072: Obtain the reference prediction value sequence according to the preliminary load prediction sequence and error correction value sequence.

[0124] Specifically, the calculation formula for the reference prediction value sequence is:

[0125] , is the preliminary load prediction sequence.

[0126] Step 2073: Use the meteorological correction prediction value sequence, the benchmark prediction value sequence, and the auxiliary feature sequence as the input of the gradient boosting tree model, so that the gradient boosting tree model outputs the first load prediction sequence.

[0127] Specifically, the gradient boosting tree model is an XGBoost (eXtreme Gradient Boosting) model; the auxiliary feature sequence includes the key meteorological elements in the current period, the key meteorological elements in the future period, the actual load error sequence in the previous prediction cycle, and the time index feature in the future period. The XGBoost gradient boosting tree model includes multiple regression trees.

[0128] Specifically, the XGBoost gradient boosting tree model aims to minimize the prediction error The objective function of the XGBoost gradient boosting tree model is:

[0129] ; where The input of the loss function is , is the result calculated at time, y i is the true load value of the i-th sample, is the predicted load value of the i-th sample; is the regularization term, f k represents the k-th regression tree, T is the total number of regression trees. In this embodiment, by minimizing ， XGBoost ensures fitting the training data and prevents the model from being too complex and causing overfitting. During the training process, XGBoost uses the second-order gradient boosting strategy to gradually construct decision trees to reduce the error. The generated model is essentially an additive combination of multiple regression trees, that is The calculation formula of is: , represents the predicted value of the k-th regression tree for the input feature vector x of the i-th sample i .

[0130] As Figure 4 shown, it is a schematic flow chart of the XGBoost gradient boosting tree model obtaining the first load prediction sequence. The way for the XGBoost gradient boosting tree model to obtain the first load prediction sequence includes the following steps:

[0131] Step 207a: Obtain the bias term, the first fusion weight of the benchmark prediction value sequence, and the second fusion weight of the meteorological correction prediction value sequence according to the auxiliary feature sequence.

[0132] Step 207b: Fuse the benchmark prediction value sequence and the meteorological correction prediction value sequence according to the first fusion weight, the second fusion weight, and the bias term to obtain the first load prediction sequence.

[0133] The calculation formula for the first load prediction sequence is: ; where represents the benchmark prediction value ， represents the meteorological correction prediction value sequence, is the bias term.

[0134] In order to further improve the robustness of the prediction results and make full use of the advantages of different models, in this embodiment, after obtaining the meteorological correction value sequence and the error correction value sequence, an XGBoost fusion mechanism is adopted, specifically the XGBoost gradient boosting tree model. The XGBoost gradient boosting tree model performs weighted fusion on the prediction results from multiple sources, namely the meteorological correction value sequence and the error correction value sequence. As an ensemble learning method, XGBoost can automatically adjust the weights and bias corrections of different predictions according to historical errors to achieve the optimal combination of prediction results.

[0135] Specifically, the input features of the XGBoost fusion model include the prediction results and relevant auxiliary information from the above-mentioned modules. The main features input to XGBoost include: the meteorological correction prediction value sequence, the benchmark prediction value sequence, and other auxiliary features. The auxiliary features can help the tree model determine how to weight different predictions in special weather, provide the actual error of the previous rolling period for correction, and distinguish the importance of different models during the day and at night.

[0136] The XGBoost fusion model of this embodiment includes the meteorological correction prediction value sequence and the benchmark prediction value sequence. Through XGBoost, two sets of prediction results can be obtained, namely the meteorological correction prediction value sequence (corrected based on the MSCNN network model) and the benchmark prediction value sequence (corrected based on the ResNet residual network model). One set emphasizes the time series pattern, and the other set focuses on meteorological factors. At the same time, the auxiliary features provide context, enabling XGBoost to dynamically adjust the trust in the two sets of results according to the environment. For example, in stable weather, the benchmark prediction value sequence may be mainly trusted, while in extreme weather, more reliance may be placed on the meteorological correction prediction value sequence.

[0137] In extreme cases, if XGBoost determines that a model's modified structure is particularly reliable in the current context, it will assign a weight close to 1 to that model's prediction and another weight close to 0. Alternatively, it may take a weighted average of the two in general. Because XGBoost is a tree model, these weights are actually fixed within each decision leaf node, corresponding to a specific rule region. However, overall, they appear to be piecewise constant functions of the input space, reflecting a rule-based weighted fusion.

[0138] In practical applications, the baseline forecast value sequence obtained from each rolling forecast will be and meteorological correction forecast series Equal input XGBoost, output final corrected fusion prediction Through offline training, XGBoost has learned how to combine the strengths of both forecasts, for example, by smoothing out potential underestimation peaks in the baseline forecast series and correcting for over-correction in the meteorologically corrected forecast series during calm weather. The fused results demonstrate lower overall error and greater stability across a variety of scenarios, meeting the forecast reliability requirements of virtual power plants and microgrids.

[0139] Step 208 : obtaining a first load sequence having the highest similarity to the load sequence in the historical load sequence, and obtaining an actual load sequence and an actual meteorological sequence in a target period after the period of the first load sequence.

[0140] Step 209: Calculate the meteorological difference sequence between the actual meteorological sequence and the predicted meteorological sequence.

[0141] Step 210: Obtain a load difference sequence according to the meteorological difference sequence.

[0142] Specifically, the above step 204 has obtained the predicted meteorological sequence corresponding to the future period after the current period. Therefore, the above-obtained predicted meteorological sequence can be directly used to calculate the meteorological difference sequence between the actual meteorological sequence and the predicted meteorological sequence, and obtain the load difference sequence based on the meteorological difference sequence.

[0143] Step 211 : superimpose the load difference sequence onto the actual load sequence to obtain a second load forecast sequence corresponding to a future period.

[0144] Step 201, step 202, step 203, step 208, step 209, step 210, and step 211 of this embodiment are substantially the same as step 101, step 102, step 103, step 105, step 106, step 107, and step 108 of the above embodiment, and will not be described again here to avoid repetition.

[0145] Rolling prediction enables the prediction to be refreshed accordingly whenever new information becomes available. For example, if the actual load trend deviates from the previous cycle's prediction, a new similarity search will select a more matching new segment, and the prediction will correct the trajectory. This continuous update mechanism greatly improves the sensitivity of the prediction to real-time changes and can promptly reflect sudden fluctuations or trend inflection points in the load. In implementation, this process can be automated as a background service, triggering calculations every 15 minutes, thereby forming a continuous rolling prediction stream.

[0146] On the other hand, this application provides an energy storage system, including: an energy storage device, an energy storage bidirectional converter, and an energy management module. The energy management module is used to implement the power load prediction method of the energy storage system as described above, and the energy management module is also used to adjust the charge and discharge strategy of the energy storage system according to the first load prediction sequence or the second load prediction sequence.

[0147] Those of ordinary skill in the art can understand that the above embodiments are specific embodiments for implementing this application, and in practical applications, various changes can be made in form and details without departing from the spirit and scope of this application.

Claims

1. A method for predicting the power load of an energy storage system, characterized in that, Including: Obtaining the load sequence of the energy storage system for the current period, and obtaining the meteorological sequence for the current period; The meteorological sequence for the current period is a sequence composed of meteorological data corresponding to multiple moments, and the meteorological data at least includes: temperature, humidity, air pressure, weather type; wherein, the weather type is represented in the form of numerical coding; Judging whether there is a load sequence in the historical load sequence of the energy storage system whose similarity to the load sequence exceeds a preset threshold; If not, according to the load sequence and the meteorological sequence, obtaining a preliminary load prediction sequence for a future period after the current period; Correcting the preliminary load prediction sequence to obtain a first corrected load prediction sequence; If so, obtaining a first load sequence with the highest similarity to the load sequence in the historical load sequence, and obtaining the actual load sequence and the actual meteorological sequence for a target period after the period where the first load sequence is located; Obtaining the predicted meteorological sequence corresponding to the future period after the current period; calculating the meteorological difference sequence between the actual meteorological sequence and the predicted meteorological sequence; obtaining the load difference sequence according to the meteorological difference sequence; Superimposing the load difference sequence on the actual load sequence to obtain a second load prediction sequence corresponding to the future period.

2. The power load prediction method for the energy storage system according to claim 1, wherein The obtaining of the preliminary load prediction sequence for the future period after the current period according to the load sequence and the meteorological sequence includes: Using the load sequence and the meteorological sequence together as the input sequence of the prediction model, so that the prediction model outputs the preliminary load prediction sequence for the future period; Wherein, the prediction model includes an encoder and a decoder, the encoder includes a plurality of stacked encoder layers, a residual connection module and a layer normalization module arranged after the stacked encoder layers, and each encoder layer includes a multi-head self-attention mechanism module and a first feed-forward neural network module; the decoder includes a masked self-attention mechanism module, a frequency domain enhancement module, an attention mechanism module, and a second feed-forward neural network module connected in sequence; The frequency domain enhancement module performs moving average decomposition on the load sequence to obtain a low-frequency trend part and a high-frequency trend part, and applies discrete Fourier transform or wavelet transform to randomly select a fixed number of frequency components as the basic frequencies for calculation by the attention mechanism module.

3. The power load prediction method for the energy storage system according to claim 2, wherein The training process of the prediction model includes: Obtaining an initial training sequence, the initial training sequence includes the feature data corresponding to the most recent N moments, and the feature data corresponding to each moment includes the load data corresponding to the moment and the meteorological data corresponding to the moment; N is an integer greater than 0; Normalize the initial training sequence to obtain a target training sequence. The formula for the normalization process is as follows: ; where x is a feature data in the initial training sequence, min(X) is the minimum feature data of the initial training sequence, max(X) is the maximum feature data of the initial training sequence, and x' is a feature data of the target training sequence after normalization; Inputting the target training sequence into the prediction model for training to obtain the trained prediction model.

4. The method for predicting the power load of the energy storage system according to claim 2, wherein, The using of the load sequence and the meteorological sequence together as the input sequence of the prediction model, so that the prediction model outputs the preliminary load prediction sequence for the future period, includes: The encoder receives the input sequence, and the length of the input sequence is M, and M is an integer greater than 0; The encoder extracts the input sequence layer by layer to obtain the feature representation of the input sequence ; The formula of the encoder is as follows: ; X is a feature data in the input sequence, and one feature data is composed of load data and meteorological data at the same moment represents the mapping result of the encoder for the input feature data X The decoder generates the preliminary load prediction sequence for the future time period according to the feature representation Zenc and the feature data X of the prediction starting point of the future time period feat ; the formula of the decoder is: ; Y is the preliminary load prediction sequence, indicating the mapping result of the decoder for the feature representation Zenc and the feature data X of the prediction starting point of the future time period feat .

5. The method for predicting the power load of the energy storage system according to claim 3, wherein The multi-head self-attention mechanism module is parallelly provided with h independent attention heads, and each of the attention heads has its own first projection matrix W i Q , W i K , W i V , where h is an integer greater than 1, and the first projection matrices W i Q , W i K , W i V are respectively used for linearly transforming queries, keys, and values; Each of the attention heads maps the input sequence according to its respective first projection matrix \(W\) i Q 、\(W\) i K 、\(W\) i V to obtain an attention output \(Q\) i ; concatenate the \(h\) attention outputs \(Q\) i in the feature dimension to obtain a concatenated output \(Concat(O_1, O_2, \ldots, O\) h ); map the concatenated output \(Concat(O_1, O_2, \ldots, O\) O back to the original dimension through a second projection matrix \(W\) h ), where the formula for the original dimension is: ; ; Q is the query vector, K is the key vector, and V is the value vector; The first feedforward neural network module outputs a mapping F(x) corresponding to each feature data x in the initial training sequence according to the original dimension ; The residual connection module obtains a residual connection output according to the mapping F(x). , the residual connection output has the following calculation formula: ; The layer normalization module performs layer normalization on the output of the residual connection to obtain , which represents the result of performing layer normalization on ; The layer normalization module is further configured to normalize the input of each encoder layer to zero mean and unit variance.

6. The power load prediction method for the energy storage system according to claim 2, wherein The masked self-attention mechanism module includes a masking matrix M. For the query vector and key vector of the decoder, when j ≤ i, M ij = 0, indicating that the position i corresponding to the key vector pays attention to the information at the position j corresponding to the query vector. When j > i, M ij = -∞, indicating that the position i corresponding to the key vector cannot pay attention to the information at the position j corresponding to the query vector. The calculation formula of the masked self-attention mechanism module is: ； Among them, represents the output result when the query vector, key vector, and value vector input to the masked self-attention mechanism module are Q, K, and V in sequence, QK T represents the dot product of vector Q and vector K, d k is the dimension of vector K, used to scale the result of QK T represents the value of the Softmax function when taking as the input.

7. The power load prediction method for the energy storage system according to claim 1, wherein Before correcting the preliminary load prediction sequence to obtain the corrected first load prediction sequence, it includes: Obtaining the predicted meteorological sequence corresponding to the future time period; Estimating the influence deviation of the predicted meteorological sequence on the preliminary load prediction sequence to obtain a meteorological correction value sequence; Obtaining an error correction value sequence corresponding to the preliminary load prediction sequence according to the load sequence and the meteorological sequence; Correcting the preliminary load prediction sequence to obtain the corrected first load prediction sequence, including: Obtaining the corrected first load prediction sequence according to the preliminary load prediction sequence, the meteorological correction value sequence, and the error correction value sequence.

8. The method for predicting the power load of the energy storage system according to claim 7, wherein, Estimating the influence deviation of the predicted meteorological sequence on the preliminary load prediction sequence to obtain a meteorological correction value sequence, including: Using the predicted meteorological sequence as the input of the meteorological correction model so that the meteorological correction model outputs the meteorological correction value sequence; Wherein, the meteorological correction model includes a plurality of parallel convolutional kernels, and the meteorological correction model further includes a non-linear activation layer and a pooling layer; the width and dilation rate of each convolutional kernel are different; The way for the meteorological correction model to obtain the meteorological correction value sequence includes: each of the convolutional kernels extracts features from each of the meteorological data in the predicted meteorological sequence to obtain the initial meteorological feature f j ; The initial meteorological feature f j is obtained through the following expression: ; represents the convolution operation on when the convolutional kernel width is k and the dilation rate is d, is the predicted meteorological sequence, W(t:t+H)=[w1,w2,...,w H , H is the prediction step, and w1, w2,..., w H are the meteorological data at different times respectively; the meteorological correction model also fuses the initial meteorological features f j output by each of the convolutional kernels to obtain the comprehensive meteorological feature corresponding to each of the meteorological data ; The comprehensive meteorological feature refines the meteorological correction value sequence of length H through the non-linear activation layer and the pooling layer , and each element in ΔYmet is the meteorological correction value corresponding to each of the meteorological data.

9. The power load prediction method for the energy storage system according to claim 8, characterized in that Obtaining an error correction value sequence corresponding to the preliminary load prediction sequence according to the load sequence and the meteorological sequence, including: Using the load sequence and the meteorological sequence as the input sequence of the error learning model so that the error learning model outputs the error correction value sequence; The loss function of the error learning model is as follows: , where \(e(t)\) is the sequence of historical error correction values learned by the error learning model, \(y(t)\) is the sequence of historical true load, is the sequence of historical predicted load.

10. The method for predicting the power load of the energy storage system according to claim 9, wherein, The error learning model is a deep residual network model; the deep residual network model includes L fully connected layers, where L is an integer greater than 1; each of the fully connected layers learns and corrects the error part not captured by the previous fully connected layer, and the output of the -th fully connected layer of the deep residual network model satisfies the following formula: , is greater than 0 and less than L; Among them, represents the input feature of the th fully connected layer at time t + i. The input feature of the th fully connected layer includes the output of the previous fully connected layer. The input feature of the first fully connected layer is the input sequence of the error learning model. is the parameter of the th fully connected layer. is the correction term mapped based on and . The calculation formula for the error correction value sequence output by the Lth fully connected layer is: .

11. The method for predicting the power load of the energy storage system according to claim 10, wherein, Obtaining the corrected first load prediction sequence according to the preliminary load prediction sequence, the meteorological correction value sequence, and the error correction value sequence, including: According to the preliminary load prediction sequence and the meteorological correction value sequence obtain a meteorological correction prediction value sequence; according to the preliminary load prediction sequence and the error correction value sequence obtain a reference prediction value sequence , the calculation formula of the reference prediction value sequence is: , is the preliminary load prediction sequence; Using the meteorological correction prediction value sequence, the reference prediction value sequence, and the auxiliary feature sequence as the input of the XGBoost gradient boosting tree model so that the XGBoost gradient boosting tree model outputs the first load prediction sequence; the auxiliary feature sequence includes the key meteorological elements of the current time period, the key meteorological elements of the future time period, the actual load error sequence of the previous prediction cycle, and the time index feature of the future time period; the XGBoost gradient boosting tree model includes multiple regression trees; The XGBoost gradient boosting tree model aims to minimize the prediction error The objective function of the XGBoost gradient boosting tree model is as follows: ； Among them, is the input of the loss function as , is the result calculated when i is the true load value of the i-th sample, is the predicted load value of the i-th sample; is the regularization term, f k represents the k-th regression tree, T is the total number of the regression trees, The calculation formula of is: represents the predicted value of the k-th regression tree for the input feature vector x i of the i-th sample.

12. The method for predicting the power load of the energy storage system according to claim 11, wherein, The manner in which the XGBoost gradient boosting tree model obtains the first load prediction sequence includes: The XGBoost gradient boosting tree model obtains a bias term according to the auxiliary feature sequence and a first fusion weight of the benchmark prediction value sequence and a second fusion weight of the meteorological correction prediction value sequence ; Fuse the reference prediction value sequence and the meteorological correction prediction value sequence according to the first fusion weight , the second fusion weight , and the bias term to obtain the first load prediction sequence. The calculation formula for the first load prediction sequence is: ; where represents the reference prediction value ， represents the meteorological correction prediction value sequence.

13. The power load forecasting method for the energy storage system according to claim 1, characterized in that, Calculating the similarity of the load sequence in the historical load sequence using the Gaussian similarity function, and the Gaussian similarity function is: ; wherein, is a scale parameter set according to the historical fluctuation amplitude, is from the time point -M + 1 to the historical load sequence, is the load sequence from the time point t - M + 1 to t, , is the load data corresponding to a moment in the load sequence.

14. The method for predicting the power load of the energy storage system according to claim 1 or 13, characterized in that, Calculating a meteorological difference sequence between the actual meteorological sequence and the predicted meteorological sequence The calculation formula is as follows: ; wherein, represents the actual meteorological sequence, represents the predicted meteorological sequence; Obtaining the load difference sequence according to the meteorological difference sequence, including: Input the meteorological difference sequence into a meteorological correction model or an error learning model to estimate the load difference sequence caused by the meteorological difference .

15. An energy storage system, characterized in that, Including: An energy storage device, an energy storage bidirectional converter, and an energy management module. The energy management module is configured to implement the power load prediction method of the energy storage system according to any one of claims 1-14. The energy management module is further configured to adjust the charge and discharge strategy of the energy storage system according to the first load prediction sequence or the second load prediction sequence.

Images