Power load prediction method, electronic equipment and storage medium

Abstract

The invention discloses a power load prediction method, electronic equipment and a storage medium. The method relates to the field of electric power systems, and comprises the following steps: obtaining an original load data sequence of an electric power system and an influence data sequence of key influence factors in a historical time period, the key influence factors being used for representing factors which can influence the load data sequence of the electric power system; the original load data sequence is subjected to iterative decomposition, at least one load sub-data sequence is obtained, and different iteration rounds correspond to different load sub-data sequences; and inputting the at least one load sub-data sequence and the influence data sequence into a load prediction model set, and performing coincidence prediction on the power system by using the load prediction model set to obtain a target load data sequence of the power system in a future time period. According to the invention, the technical problem of low accuracy of power load prediction in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the field of power systems, and in particular, to a power load forecasting method, an electronic device, and a storage medium. Background Art

[0002] With the continuous growth of global energy demand and the rapid development of renewable energy technologies, power load forecasting plays a crucial role in power system planning, operation scheduling, and market transactions. Accurate load forecasting can help grid managers optimize the allocation of power resources, reduce the operating costs of power systems, enhance the stability and reliability of power systems, and is also of great significance for formulating competition strategies in the power market.

[0003] In recent years, machine learning and deep learning technologies have been widely applied in the field of load forecasting due to their powerful non-linear modeling capabilities and adaptive learning mechanisms. These technologies can learn complex patterns from historical data and capture the deep features of time series through a multi-layer neural network structure. However, it is relatively difficult to select and adjust hyperparameters during the training process of prediction models based on machine learning and deep learning technologies, and overfitting or underfitting phenomena are likely to occur, affecting the accuracy of prediction results.

[0004] To address the above problems, no effective solutions have been proposed yet. Summary of the Invention

[0005] Embodiments of the present invention provide a power load forecasting method, an electronic device, and a storage medium to at least solve the technical problem of low accuracy in power load forecasting in related technologies.

[0006] According to one aspect of the embodiments of the present invention, a power load forecasting method is provided, including: obtaining, within a historical time period, an original load data sequence of a power system and an influence data sequence of key influencing factors, where the key influencing factors are used to characterize factors that will affect the load data sequence of the power system; performing iterative decomposition on the original load data sequence to obtain at least one load sub-data sequence, where the load sub-data sequences corresponding to different iterative rounds are different; inputting the at least one load sub-data sequence and the influence data sequence into a load forecasting model set, and using the load forecasting model set to perform compliance forecasting on the power system to obtain a target load data sequence of the power system within a future time period.

[0007] Further, iteratively decompose the original load data sequence to obtain at least one load sub-data sequence, including: for any one of multiple iterative rounds, obtain the noise signal and the residual load data sequence corresponding to the iterative round, where the residual load data sequence is used to represent the load data sequence in the original load data sequence except for the load sub-data sequence corresponding to the historical iterative rounds, and the residual load data sequence in the first iterative round is the original load data sequence, and the historical iterative rounds are the iterative rounds executed before the iterative round in the multiple iterative rounds; add the noise signal to the residual load data sequence to obtain a noise load data sequence; perform signal decomposition on the noise load data sequence to obtain the load sub-data sequence corresponding to the iterative round.

[0008] Further, obtaining the noise signal corresponding to the iterative round includes: obtaining the noise component and the signal-to-noise ratio corresponding to the iterative round, and the standard deviation of the residual load data sequence; obtaining a noise adjustment coefficient based on the ratio of the signal-to-noise ratio to the standard deviation; adjusting the noise component based on the noise adjustment coefficient to obtain the noise signal.

[0009] Further, the load prediction model set includes: a recurrent neural network model and a deep convolutional network model; input at least one load sub-data sequence and the impact data sequence into the load prediction model set, and use the load prediction model set to perform load prediction on the power system to obtain the target load data sequence of the power system in the future time period, including: input at least one load sub-data sequence and the impact data sequence into the recurrent neural network model to obtain a first load data sequence; input at least one load sub-data sequence and the impact data sequence into the deep convolutional network model to obtain a second load data sequence; perform weighted processing on the first load data sequence and the second load data sequence based on the first model weight corresponding to the recurrent neural network model and the second model weight corresponding to the deep convolutional network model to obtain the target load data sequence.

[0010] Further, the method further includes: obtaining a training data sequence for a first time period and a load data sequence for a second time period, where the end time of the first time period is before the start time of the second time period, the training data sequence includes: the load data sequence of the power system in the first time period and the data sequence of key influencing factors in the first time period, and the load data sequence includes: the load data sequence of the power system in the first time period; based on the training data sequence, using a recurrent neural network model to predict the load data sequence of the power system in the second time period to obtain a first prediction sequence, and based on the training data sequence, using a deep convolutional network model to predict the load data sequence of the power system in the second time period to obtain a second prediction sequence; performing weighted processing on the first prediction sequence and the second prediction sequence based on initial weights, where the initial weights include: a first initial weight corresponding to the recurrent neural network model and a second initial weight corresponding to the deep convolutional network model; matching the load prediction sequence with the load data sequence to obtain a sequence matching result; adjusting the initial weights based on the sequence matching result to obtain a first model weight and a second model weight.

[0011] Further, adjusting the first initial weight and the second initial weight based on the sequence matching result to obtain a first model weight and a second model weight includes: in response to the sequence matching result indicating that the load prediction sequence does not match the load data sequence, obtaining the current iteration round, and the first return value of multiple weight adjustment actions in the weight adjustment action set in the current iteration round, where the first return value is used to represent the selection probability of different weight adjustment actions; based on the current iteration round and the first return value, selecting a target adjustment action from the weight adjustment action set; adjusting the initial weights based on the target adjustment action to obtain a first adjusted weight and a second adjusted weight; based on the first adjusted weight and the second adjusted weight, re-determining the load prediction sequence, and when the new load prediction sequence matches the load data sequence, determining the first adjusted weight as the first model weight and determining the second adjusted weight as the second model weight.

[0012] Further, obtaining the first return value of multiple weight adjustment actions in the weight adjustment action set in the current iteration round includes: obtaining the second return value of the target adjustment action selected in the previous iteration round, and the load matching result in the previous iteration round; adjusting the current return value of multiple weight adjustment actions based on the second return value and the load matching result to obtain the first return value.

[0013] Further, the method further includes: obtaining multiple data sequences of different initial influencing factors and multiple load data sequences of the power system in multiple time periods, where the initial influencing factors include key influencing factors; determining multiple mutual information values of different initial influencing factors based on the multiple data sequences and the multiple load data sequences, where different mutual information values correspond to different data sequences and different load data sequences; performing normalization processing on the multiple mutual information values to obtain maximum information coefficients corresponding to different initial influencing factors; and selecting key influencing factors from the initial influencing factors based on the maximum information coefficients, where the maximum information coefficient of the key influencing factors is greater than the maximum information coefficients of other influencing factors, and the other influencing factors are used to represent the influencing factors other than the key influencing factors among the initial influencing factors.

[0014] According to another aspect of the embodiments of the present invention, there is also provided a power load prediction device, including: a first acquisition module, configured to acquire an original load data sequence of the power system and an influence data sequence of key influencing factors in a historical time period, where the key influencing factors are used to represent factors that will affect the load data sequence of the power system; a first decomposition module, configured to perform iterative decomposition on the original load data sequence to obtain at least one load sub-data sequence, where the load sub-data sequences corresponding to different iterative rounds are different; and a first prediction module, configured to input the at least one load sub-data sequence and the influence data sequence into a load prediction model set, and use the load prediction model set to perform compliance prediction on the power system to obtain a target load data sequence of the power system in a future time period.

[0015] According to another aspect of the embodiments of the present invention, there is also provided an electronic device, including: a memory storing an executable program; and a processor configured to run the program, where when the program runs, it executes the methods in the various embodiments of the present invention.

[0016] According to another aspect of the embodiments of the present invention, there is also provided a computer-readable storage medium, where the computer-readable storage medium includes a stored executable program, and when the executable program runs, it controls the device where the computer-readable storage medium is located to execute the methods in the various embodiments of the present invention.

[0017] According to another aspect of the embodiments of the present invention, there is also provided a computer program product, including a computer program, where the computer program implements the methods in the various embodiments of the present invention when executed by a processor.

[0018] According to another aspect of the embodiments of the present invention, there is also provided a computer program product, including a non-volatile computer-readable storage medium, where the non-volatile computer-readable storage medium stores a computer program, and the computer program implements the methods in the various embodiments of the present invention when executed by a processor.

[0019] According to another aspect of the embodiments of the present invention, there is also provided a computer program which, when executed by a processor, implements the methods in the various embodiments of the present invention.

[0020] In the embodiments of the present invention, the following steps are adopted: obtaining the original load data sequence of the power system and the influence data sequence of key influencing factors within a historical time period; performing iterative decomposition on the original load data sequence to obtain at least one load sub-data sequence; inputting the at least one load sub-data sequence and the influence data sequence into a load prediction model set, and using the load prediction model set to perform load prediction on the power system to obtain the target load data sequence of the power system within a future time period. By decomposing the original load data sequence, the complex load data sequence is decomposed into a series of simpler and more physically meaningful load sub-data sequences, so as to reduce the prediction difficulty. Subsequently, the at least one load sub-data sequence obtained by iterative decomposition, together with the influence data sequence of key influencing factors, is input into a load prediction model set containing multiple prediction models to combine the advantages of different prediction models and predict the decomposed load sub-data sequences. Finally, the prediction results of the load prediction model set are weighted and integrated to obtain the final target load data sequence, so as to utilize the complementarity between different models to improve the overall prediction accuracy, achieve the purpose of reducing the power load prediction error, thereby realizing the technical effect of improving the accuracy of power load prediction, and further solving the technical problem of low accuracy of power load prediction in the related art. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention. In the drawings:

[0022] Figure 1 is a flowchart of a power load prediction method according to an embodiment of the present invention;

[0023] Figure 2 is a schematic diagram of the network structure of an optional BiGRU model according to an embodiment of the present invention;

[0024] Figure 3 is a schematic diagram of the structure of an optional deep residual graph convolutional model according to an embodiment of the present invention;

[0025] Figure 4 is a flowchart of the load prediction process of an optional deep learning hybrid model according to an embodiment of the present invention;

[0026] Figure 5 is a schematic diagram of a power load prediction device according to an embodiment of the present invention. Detailed implementation manners

[0027] In order to enable those skilled in the art of the present technology to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0028] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily need to be used to describe a specific order or sequence. It should be understood that such data used can be interchanged under appropriate circumstances so that the embodiments of the present invention described herein can be implemented in an order different from those illustrated or described herein. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device comprising a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0029] According to an embodiment of the present invention, an embodiment of a power load forecasting method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. And although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that here.

[0030] Figure 1 is a flowchart of a power load forecasting method according to an embodiment of the present invention, as Figure 1 shown, the method includes the following steps:

[0031] Step S102, obtaining, within a historical time period, the original load data sequence of the power system and the influence data sequence of key influencing factors, where the key influencing factors are used to characterize the factors that will affect the load data sequence of the power system.

[0032] The above historical time period may refer to a past time interval used for training the prediction model. The selection of the above historical time period can be based on the availability of historical data of power load, periodicity (such as daily, weekly, monthly cycles) and the amount of data required for model training. For example, if the power load for the next day is to be predicted, the historical time period can cover the load data of the recent few days, weeks or months to capture short-term and long-term load change trends, but not limited thereto.

[0033] The above-mentioned original load data sequence may refer to the time series of the actual load data of the power system recorded during the above-mentioned historical time period. The above-mentioned original load data sequence is the basis for training the prediction model and directly reflects the natural variation law of the power load over time.

[0034] The above-mentioned key influencing factors may refer to the factors that can significantly affect the power load except for time itself. The above-mentioned key influencing factors can be selected based on historical data analysis and domain knowledge, so as to focus on the variables that have a significant impact on the power load. For example, the above-mentioned key influencing factors may be temperature, humidity, wind speed, etc., but are not limited thereto.

[0035] The above-mentioned influence data sequence may refer to the data records of the key influencing factors during the historical time period. The above-mentioned influence data sequence can be used together with the original load data sequence to analyze the relationship between the load and the influencing factors.

[0036] In an alternative embodiment, considering that the original load data sequence of the power system records the actual situation of the power demand changing over time, it can provide a data basis for the load prediction system (hereinafter referred to as the prediction system). The power company or grid operator can collect the power load data of the power system every hour or at a finer time interval (such as 15 minutes) through the power monitoring system or smart meters. The collection of these power load data can cover multiple historical time periods, such as every day in the past year, so that the original load data sequence constructed by the prediction system based on the above-mentioned power load data can reflect the influence of different seasons, different weather conditions and various social and economic activities on the power load. In addition, the prediction system can also identify the factors that have a significant impact on the power load based on domain knowledge and historical data analysis as the above-mentioned key influencing factors, and further construct the influence data sequence of the key influencing factors based on the key influencing factors. Through the comprehensive analysis of the original load data sequence and the influence data sequence, the prediction system can more accurately capture the law of the power load changing over time, and at the same time consider the dynamic influence of external factors, so as to improve the accuracy and stability of the load prediction.

[0037] Step S104: Perform iterative decomposition on the original load data sequence to obtain at least one load sub-data sequence, where the load sub-data sequences corresponding to different iteration rounds are different.

[0038] The above-mentioned load sub-data sequence may refer to the sub-sequences obtained by decomposing the original load data sequence into multiple simpler and more distinct characteristic (such as trend, seasonality, noise, etc.) sub-sequences through data decomposition techniques. Each load sub-data sequence represents different components in the original load data sequence, such as long-term trend, seasonal fluctuation, periodic change or random noise, etc., but is not limited thereto.

[0039] In an alternative embodiment, considering that the original load data usually contains multiple components such as long-term trends, seasonal fluctuations, short-term cycles, and random noise, and these multiple components are intertwined, making load forecasting relatively complex. By iteratively decomposing the original data sequence, a series of simpler and purer load sub-data sequences can be obtained, enabling each load sub-data sequence to focus on a certain aspect of the data, such as high-frequency changes, medium-frequency fluctuations, or low-frequency trends, thereby reducing the processing difficulty of the prediction system and improving the accuracy and stability of the prediction. In this embodiment, through the above decomposition process, the complex patterns and dynamic characteristics in the original load data can be separated, enabling subsequent data analysis to focus more on the attributes of each sub-sequence, thereby improving the accuracy and robustness of the prediction.

[0040] Step S106, input at least one load sub-data sequence and the impact data sequence into the load forecasting model set, and use the load forecasting model set to perform load forecasting on the power system to obtain the target load data sequence of the power system in the future time period.

[0041] The above load forecasting model set can be a model library composed of multiple different forecasting models, and these models are each good at dealing with different aspects of power load data. For example, the above load forecasting model set can include at least one or more of the following models: the BiGRU (Bidirectional Gated Recurrent Unit) model based on deep learning, the deep residual graph convolutional model, etc., but not limited to this. The above load forecasting model set can comprehensively utilize the advantages of different algorithms to improve the accuracy and robustness of power load forecasting. The above target load data sequence can refer to the prediction result of the load demand of the power system in the future time period.

[0042] In an alternative embodiment, a model set composed of different types of forecasting models can make full use of their respective advantages. For example, the BiGRU model can capture long-term dependencies when processing time series data and is suitable for predicting long-term trends and periodic patterns; the deep residual graph convolutional model can process data with spatial correlation and complex local features and is more effective for predicting the mutual influence between different nodes in the power grid and local load changes. Since the above load forecasting model set can make up for the deficiencies of a single model, thereby improving the accuracy of load forecasting, based on this, by inputting at least one load sub-data sequence and the impact data sequence into the load forecasting model set and using the load forecasting model set to perform load forecasting on the power system, the prediction system can combine the prediction results of the above different models to obtain a more comprehensive and accurate target load data sequence.

[0043] For ease of understanding, Figure 2It is a schematic diagram of the network structure of an optional BiGRU model according to an embodiment of the present invention, as Figure 2 shown. This structure includes an input x t , four sigmoid activation functions (denoted by σ), two tanh (Hyperbolic Tangent) functions, and two update gates u t , and two reset gates g t . Among them, the input x t can be any feature related to prediction, such as the power load, meteorological data, or other influencing factors in a time series. The sigmoid activation function can map any real value to between 0 and 1 and is used in the gating mechanism to determine the degree of information transmission. The tanh function is used for the update of the hidden state and is used in combination with the output of the sigmoid function to adjust the information transmission. The update gate uses the sigmoid activation function, and its output is a value between 0 and 1, which is used to control how much historical information and how much new input information should be retained in the cell state at the current moment. The reset gate also uses the sigmoid activation function, and its role is to determine how to mix the current input and historical information when updating the cell state. The GRU (Gated Recurrent Unit) consists of the update gate u t and the reset gate g t , and is specifically expressed as shown in the following formula:

[0044]

[0045] In the formula, x t is the input, σ is the sigmoid activation function, represents the weight matrix of the update gate, represents the weight matrix of the reset gate, is the weight, h t is the state of the hidden layer at time t, and h t-1 is the state of the hidden layer at time t - 1, is the candidate hidden state at time t. The BiGRU obtains the final prediction value by integrating the prediction values of the forward GRU and the backward GRU. The specific calculation process can be shown as the following formula:

[0046]

[0047] In the formula, is the output weight of the forward GRU, is the output weight of the backward GRU, is the output of the forward GRU, is the output of the backward GRU.

[0048] Figure 3is a schematic diagram of the structure of an optional deep residual graph convolution model according to an embodiment of the present invention, such as Figure 3 As shown in the figure, the structure includes an input layer (Input data), three GCN (Graph Convolutional Network) layers, three batch normalization layers (Batch Normalization, BN), two convolutional layers (Conv), one maximum pooling layer (Maximum pooling), three dense layers (Dense), one dropout layer (Dropout), and an output layer (Output). Among them, the data of the input layer is first sent to the graph convolutional network layer. The graph convolutional network can process graph structure data and extract the dependency relationship between nodes and the characteristics of the graph. Each graph convolutional network layer is used to extract deeper features from the input data. The graph convolution operation takes the adjacency matrix and node feature matrix of the graph as input, and captures the relationship between nodes and the structural information of the graph through specific convolution operations. After each graph convolutional network layer, there is a batch normalization layer to accelerate the training process and improve the generalization ability of the model. After the graph convolutional network layer, there are two convolutional layers, which are used to further extract local features or patterns in time series data. After the convolutional layer, there is a maximum pooling layer, which is used to strengthen the detection of features and reduce the amount of calculation and parameters of subsequent layers. After the maximum pooling layer, there are two dense layers, which are used for more complex and deeper feature learning. After the two dense layers, there is a dropout layer, which is used to randomly delete the output of some neurons during training to prevent overfitting. After the dropout layer, there is a dense layer, followed by an output layer, which is used to generate the final prediction results of the model. The structure of the output layer and the choice of activation function depend on the specific application scenario and prediction target.

[0049] Specifically, GCN can fully integrate node information and embed data features according to the information of adjacent nodes to explore the potential connections between features. Specifically, the input of GCN includes the node feature matrix φ and the adjacency matrix φ is the feature set of all nodes in the graph network, It is the correlation between all nodes in the graph network. The representation of each layer of GCN can be shown as follows:

[0050] H (C+1) =f(H (C) ,φ).

[0051] Among them, H (C) is the output of layer C, H (C+1)is the output of the (C + 1)-th layer. The GCN uses a convolutional operation to add self-loops to the adjacency matrix, obtaining a new adjacency matrix and normalizing it. The calculation formula of the GCN can be shown as follows:

[0052]

[0053] where I is the identity matrix, Θ is a diagonal matrix, the diagonal elements of which are the sum of each row of (C) , and the rest of the elements are 0. ω

[0054]

[0055] In the formula, H (1) , H (3) , H (5) are the outputs of the 1st, 3rd, and 5th layers in the GCN layer. ω (1) , ω (3) , ω (5) are the corresponding weights of the 1st, 3rd, and 5th layers. H (2) , H (4) are the outputs of the 2nd and 4th layers in the BN layer. Among them, H (2) = BN(H (1) ), H (4) = BN(H (3) ).

[0056] In an embodiment of the present invention, the original load data sequence of the power system and the influence data sequence of key influencing factors are obtained within a historical time period; the original load data sequence is iteratively decomposed to obtain at least one load sub-data sequence; the at least one load sub-data sequence and the influence data sequence are input into a load prediction model set, and the load prediction model set is used to perform compliance prediction on the power system to obtain the target load data sequence of the power system within a future time period. By decomposing the original load data sequence, the complex load data sequence is decomposed into a series of simpler and more physically meaningful load sub-data sequences to reduce the prediction difficulty. Subsequently, the at least one load sub-data sequence obtained by iterative decomposition, together with the influence data sequence of key influencing factors, is input into a load prediction model set containing multiple prediction models to combine the advantages of different prediction models and predict the decomposed load sub-data sequences. Finally, the prediction results of the load prediction model set are weighted and integrated to obtain the final target load data sequence, so as to utilize the complementarity between different models to improve the overall prediction accuracy, achieve the purpose of reducing the power load prediction error, thereby realizing the technical effect of improving the accuracy of power load prediction, and further solving the technical problem of low accuracy of power load prediction in the related art.

[0057] Further, iteratively decomposing the original load data sequence to obtain at least one load sub-data sequence includes: for any one of multiple iterative rounds, obtaining the noise signal and the residual load data sequence corresponding to the iterative round, where the residual load data sequence is used to represent the load data sequence in the original load data sequence except for the load sub-data sequence corresponding to the historical iterative round, and the residual load data sequence of the first iterative round is the original load data sequence, and the historical iterative round is the iterative round executed before the iterative round in the multiple iterative rounds; adding the noise signal to the residual load data sequence to obtain the noise load data sequence; and performing signal decomposition on the noise load data sequence to obtain the load sub-data sequence corresponding to the iterative round.

[0058] The above residual load data sequence may be the load data sequence in the original load data sequence except for the load sub-data sequence corresponding to the historical iterative round. The above residual load data sequence plays a bridging role in the iterative decomposition process, connecting the original data sequence and the decomposed sub-sequence. The above noise load data sequence may be a sequence formed by adding the residual load data sequence of the current round and the corresponding noise signal.

[0059] In an alternative embodiment, the residual load data sequence of the first round of iteration is the original load data sequence itself, which contains all the original information of the power load, such as long-term trends, seasonal fluctuations, daily cycle changes, and random noise, etc. As the algorithm iterates, a load sub-data sequence is extracted from the current residual load data sequence in each round. This load sub-data sequence contains components of a certain frequency or pattern in the residual sequence. Once this load sub-data sequence is extracted, the prediction system can remove the information contained in this load sub-data sequence from the residual sequence, and the remaining part is reconstituted into a new residual load data sequence to represent the remaining information in the original data except for the patterns that have been decomposed and extracted. The prediction system can iterate the above process until the residual load data sequence no longer contains decomposable patterns or frequency components, or until the total number of current iteration rounds reaches a preset number of iterations. In addition, in each round of iterative decomposition, the prediction system can also add a noise signal to the residual load data sequence to obtain a noise load data sequence, thereby enhancing the noise tolerance ability, reducing the pseudo-mode phenomenon in prediction, and improving the robustness of the prediction model. After obtaining the above noise load data sequence, the prediction system can use the ICEEMDAN (Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) algorithm to decompose the noise load data sequence, thereby obtaining the load sub-data sequence corresponding to the iteration round. In the above processing steps, the ICEEMDAN algorithm enhances the noise tolerance of the prediction system by introducing a noise signal, enabling the prediction system to maintain stable and accurate prediction capabilities when facing complex and changing environments.

[0060] For example, when the prediction system uses the ICEEMDAN algorithm for signal decomposition, it can first add K groups of GWN ζ (k) (k = 1, 2, …, K) load sequences with an expected value of 0 to obtain K decomposed signals x (k) which can be shown as follows:

[0061] x (k) = x + τ0E1(ζ (k) ).

[0062] In the formula, τ0 is the ratio of the signal-to-noise ratio to the standard deviation, and E i (·) is the operator of the i-th white noise component generated. Through EMD decomposition, initially i = 1. When i = 1, the prediction system can calculate the first group of residuals and obtain the first modal component as shown in the following formula:

[0063]

[0064] In the formula, r1 is the first group of residuals, Z1 is the first modal component, and N(·) is the local average value of the signal. Add GWN to r1 and use local mean decomposition for x (2) It can be shown as follows:

[0065] x (2) = r1 + τ1E2(ζ (2) ).

[0066] Subsequently, the prediction system can calculate the second group of residuals r2 and the second modal component Z2 as shown in the following formula:

[0067]

[0068] Furthermore, obtaining the noise signal corresponding to the iteration round includes: obtaining the noise component and signal-to-noise ratio corresponding to the iteration round, and the standard deviation of the residual load data sequence; obtaining the noise adjustment coefficient based on the ratio of the signal-to-noise ratio to the standard deviation; and adjusting the noise component based on the noise adjustment coefficient to obtain the noise signal.

[0069] The above-mentioned noise component can refer to the random signal artificially added to the residual load data sequence in the ICEEMDAN algorithm to enhance the stability and robustness of the decomposition. It is usually composed of a group of white noises, whose characteristics are that the mean value is zero, the variance is one, and it is uniformly distributed in the frequency spectrum.

[0070] The above-mentioned signal-to-noise ratio can be an index measuring the ratio of signal strength to noise strength. In the context of the ICEEMDAN algorithm, the signal-to-noise ratio is used to describe the strength ratio between the residual load data sequence (signal) and the added noise component.

[0071] The above-mentioned noise adjustment coefficient can be a parameter calculated based on the ratio of the signal-to-noise ratio to the standard deviation of the residual load data sequence, and is used to adjust the magnitude of the noise component to ensure that in each iteration, the relative strength of the noise and the signal is in a relatively optimal range.

[0072] In an alternative embodiment, the above noise component can be composed of a set of white noises. The characteristics of the noise component can be that the mean is zero and the variance is one. The above signal-to-noise ratio is the ratio of the signal strength to the noise strength, which is used to measure the relative magnitude of the effective information in the signal and the noise interference. In the ICEEMDAN algorithm, the noise component and the signal-to-noise ratio are the basic parameters for generating the noise signal. Therefore, the prediction system can first obtain the noise component and the signal-to-noise ratio corresponding to the iteration round. Further considering that the standard deviation, as a statistical index, can be used to measure the dispersion degree of the data set, that is, the difference size between the data points and the average value. In each iteration, the ICEEMDAN algorithm will calculate the standard deviation of the current residual load data sequence to understand the volatility and complexity of the current residual load data sequence. Therefore, the prediction system can also obtain the standard deviation of the residual load data sequence. After obtaining the noise component and the signal-to-noise ratio corresponding to the iteration round, as well as the standard deviation of the residual load data sequence, based on the ratio of the signal-to-noise ratio and the standard deviation, the prediction system can use the ICEEMDAN algorithm to calculate the noise adjustment coefficient to adjust the amplitude of the noise component to ensure that the intensity of the noise signal matches the characteristics of the residual load data sequence. Finally, the prediction system can use the adjusted noise component and the determined noise adjustment coefficient to generate a noise signal suitable for the current iteration round. This signal will be added to the residual load data sequence to form a noise load data sequence, so that when the prediction model processes the decomposed load sub-data sequence, it can focus more on the time series characteristics of the signal itself rather than being disturbed by noise or outliers, thereby improving the accuracy and reliability of the prediction.

[0073] Furthermore, the load prediction model set includes: a recurrent neural network model and a deep convolutional network model; inputting at least one load sub-data sequence and the influence data sequence into the load prediction model set, and using the load prediction model set to perform load prediction on the power system to obtain the target load data sequence of the power system in the future time period, including: inputting at least one load sub-data sequence and the influence data sequence into the recurrent neural network model to obtain the first load data sequence; inputting at least one load sub-data sequence and the influence data sequence into the deep convolutional network model to obtain the second load data sequence; based on the first model weight corresponding to the recurrent neural network model and the second model weight corresponding to the deep convolutional network model, performing weighted processing on the first load data sequence and the second load data sequence to obtain the target load data sequence.

[0074] The above first load data sequence can be the data sequence output after inputting at least one load sub - data sequence and impact data sequence into a recurrent neural network model. The above second load data sequence can be the data sequence output after inputting at least one load sub - data sequence and impact data sequence into a deep convolutional network model. The above first model weight can be the weight assigned by the recurrent neural network model when contributing to the final prediction result. The above second model weight can be the weight assigned by the deep convolutional network model when contributing to the final prediction result.

[0075] In an alternative embodiment, the load prediction model set can include a recurrent neural network model and a deep convolutional network model. These two models have their respective advantages for time - series analysis and spatial feature extraction. For example, the recurrent neural network model is suitable for capturing long - term time dependencies in data and can effectively process the historical information of sequence data, while the deep convolutional network model performs well in extracting local features and processing the spatial structure of image or sequence data. When at least one load sub - data sequence and impact data sequence, such as real - time or historical data like meteorological conditions and holiday information, are input into the above load prediction model set, the recurrent neural network model and the deep convolutional network model in the load prediction model set can make full use of their characteristics to extract the key information for load prediction.

[0076] Specifically, when the prediction system inputs at least one load sub - data sequence and impact data sequence into the recurrent neural network model, the recurrent neural network model can learn the time - series characteristics of load changes and predict the future change trend of the load, outputting the first load data sequence. When the prediction system inputs the load sub - data sequence and impact data sequence into the deep convolutional network model, the deep convolutional network model can capture the load patterns related to geographical location or weather conditions through convolution operations and spatial feature extraction capabilities, and output the second load data sequence.

[0077] After obtaining the first load data sequence and the second load data sequence, the prediction system also needs to determine the weights of the above recurrent neural network model and deep convolutional network model in the final prediction result to achieve a reasonable combination of the model prediction results. The determination of the above model weights can be achieved through intelligent optimization algorithms. For example, the IQL (Improved Q - Learning) algorithm, but not limited to this. The above IQL algorithm can dynamically adjust the first model weight and the second model weight based on the accuracy of historical predictions and the adaptability of the model to load components, thereby ensuring a better prediction combination. Based on the determined model weights, the prediction system can perform weighted processing on the above first load data sequence and second load data sequence to generate the final target load data sequence.

[0078] The above Q-learning algorithm can obtain experience and knowledge through continuous trial and error, and establish a Q-table for the state-action relationship. The iterative formula of the Q-table can be shown as follows:

[0079] Q V+1 (S V ,a V )&=

[0080] Q V (S V ,a V )+ε[R(S V ,S V+1 ,a V )+λmaxQ V (S V+1 ,a ′ )-Q V (S V ,a V )]。

[0081] In the formula, Q V (S V ,a V ) is the Q value of executing action a V in state S V , R(S V ,S V+1 ,a V ) is the reward value from state S V to S V+1 , a ′ is any action value in the action space of S V+1 , ε is the learning factor, with a value of 0.01, λ is the discount factor, with a value of 0.9, and Q V (S V+1 ,a ′ ) is the Q value of executing action a V+1 in state S ′ . In the original Q-learning algorithm, a greedy mechanism is usually used for action selection, which lacks intelligence. In this embodiment, a new action selection mechanism is proposed, which is specifically shown as follows:

[0082]

[0083] In the formula, a Qmax is the action corresponding to the maximum Q value in the current state, a random is a randomly selected action, can represent the number of random selections, set to 6, T is the current iteration number, V takes 10000, and δ is a random number, with a value range of [0,1]. In In the early stage, the action selection is random, which is a process of accumulating experience. In In the later stage, the action selection is based on the action corresponding to the maximum Q value, which is a process of exploiting experience.

[0084] Furthermore, the method further includes: obtaining a training data sequence for a first time period and a load data sequence for a second time period, where the end time of the first time period is before the start time of the second time period. The training data sequence includes: the load data sequence of the power system in the first time period and the data sequence of key influencing factors in the first time period. The load data sequence includes: the load data sequence of the power system in the first time period. Based on the training data sequence, using a recurrent neural network model to predict the load data sequence of the power system in the second time period, obtaining a first prediction sequence, and based on the training data sequence, using a deep convolutional network model to predict the load data sequence of the power system in the second time period, obtaining a second prediction sequence. Performing weighted processing on the first prediction sequence and the second prediction sequence based on initial weights, obtaining a load prediction sequence, where the initial weights include: a first initial weight corresponding to the recurrent neural network model and a second initial weight corresponding to the deep convolutional network model. Matching the load prediction sequence with the load data sequence to obtain a sequence matching result. Adjusting the initial weights based on the sequence matching result to obtain a first model weight and a second model weight.

[0085] The above training data sequence can be a data set for training the recurrent neural network model and the deep convolutional network model. The above training data sequence can be composed of the load data sequence of the power system in the above first time period and the data sequence of key influencing factors in the first time period. The above first prediction sequence can be the result of the recurrent neural network model predicting the load data sequence of the power system in the second time period based on the above training data sequence. The above second prediction sequence can be the output of the deep convolutional network model predicting the load data sequence in the second time period based on the above training data sequence.

[0086] The above initial weights can be initial setting values used to reflect the relative importance of the prediction results of different models. The above initial weights can include a first initial weight corresponding to the recurrent neural network model and a second initial weight corresponding to the deep convolutional network model. The above first initial weight can be the proportion of the prediction result of the recurrent neural network model in the final load prediction sequence set initially. The above second initial weight can be the proportion of the prediction result of the deep convolutional network model in the final load prediction sequence set initially.

[0087] The above sequence matching result can be the comparison result between the load prediction sequence and the actual load data sequence in the second time period. The above sequence matching result is an important basis for evaluating the prediction performance of different models and adjusting the weights of different models.

[0088] In an alternative embodiment, the above training data sequence includes the load data sequence of the power system in the first time period and the data sequence of key influencing factors in the first time period. These data sequences generally include the load values of the power system in the historical time period, as well as other factors closely related to load changes, such as temperature, humidity, wind speed, holiday information, the difference between weekdays and weekends, etc. Therefore, the above training data sequence can be used to train the recurrent neural network model and the deep convolutional network model. The load data sequence in the second time period is the power load in the future time period to be predicted and is the benchmark for evaluating the prediction performance of the model. The acquisition of the above training data sequence and the load data sequence in the second time period ensures that the training and prediction processes of the recurrent neural network model and the deep convolutional network model are based on the actual operation data of the power system, improving the reality and accuracy of the prediction. Using the above training data sequence, the recurrent neural network model can learn the time series characteristics of the power load, including the periodic changes of the load, long-term trends, etc., and output the first prediction sequence. The deep convolutional network model can extract local features and patterns in the power load data through convolutional layers and pooling layers, and output the second prediction sequence. After obtaining the above first prediction sequence and second prediction sequence, the prediction system also needs to determine the relative importance of the prediction results of the recurrent neural network model and the deep convolutional network model based on the initial weights, where the above initial weights include the first initial weight and the second initial weight. The prediction system can perform weighted combination on the first prediction sequence and the second prediction sequence according to the first initial weight and the second initial weight, so as to obtain the above load prediction sequence. This load prediction sequence synthesizes the prediction results of time and space characteristics, and thus can utilize the advantages of both the recurrent neural network model and the deep convolutional network model at the same time, further improving the comprehensiveness and accuracy of the prediction. Subsequently, the prediction system can match the above load prediction sequence with the actual power load data sequence in the second time period to obtain a sequence matching result, which can reflect the accuracy of the prediction. Finally, the prediction system can adjust the initial weights based on the above sequence matching result to obtain better first model weight and second model weight. The above adjustment process can adopt the IQL algorithm or other intelligent optimization algorithms, and continuously adjust the model weights through multiple rounds of iteration to improve the accuracy of the final load prediction sequence.

[0089] Further, the first initial weight and the second initial weight are adjusted based on the sequence matching result to obtain the first model weight and the second model weight, including: in response to the sequence matching result indicating that the load prediction sequence does not match the load data sequence, obtaining the current iteration round and the first return value of multiple weight adjustment actions in the weight adjustment action set at the current iteration round, where the first return value is used to characterize the selection probability of different weight adjustment actions; based on the current iteration round and the first return value, selecting a target adjustment action from the weight adjustment action set; adjusting the initial weight based on the target adjustment action to obtain the first adjusted weight and the second adjusted weight; based on the first adjusted weight and the second adjusted weight, re-determining the load prediction sequence, and when the new load prediction sequence matches the load data sequence, determining the first adjusted weight as the first model weight and determining the second adjusted weight as the second model weight.

[0090] The above weight adjustment action set can be a set of operations available for selection in each iteration round for adjusting the weights of the recurrent neural network and the deep convolutional network models. The above first return value can be an index for evaluating the effects of different weight adjustment actions, which reflects the performance change of the prediction model after executing the current weight adjustment action at the current iteration round. The calculation of the above first return value can be based on the feedback of the sequence matching result and the prediction error, that is, the difference between the load prediction sequence and the actual load data sequence. The above selection probability can refer to the probability of selecting a specified weight adjustment action from the above weight adjustment action set at the current iteration round.

[0091] In an alternative embodiment, if the sequence matching result evaluated by the prediction system using the IQL algorithm shows that the load prediction sequence does not match the load data sequence, that is, the prediction error is large, it means that the current configuration of the prediction system fails to fully capture the dynamic changes and key influencing factors of the power load, and it is necessary to adjust the weights of the recurrent neural network and the deep convolutional network models to improve the prediction performance. At this time, the prediction system can trigger the weight adjustment process. Since in each iteration, the IQL algorithm used by the prediction system records the current iteration round and the first return value corresponding to each action in the weight adjustment action set at this iteration round, the prediction system can relatively easily obtain the current iteration round and the first return values of multiple weight adjustment actions in the weight adjustment action set at the current iteration round. The above first return value is an evaluation index, which is based on the performance change of the prediction model after performing a specific weight adjustment action and is used to characterize the effects of different weight adjustment actions. The first return value can not only be used to calculate the direct effect of the current adjustment action, but also affect the selection probability of weight adjustment actions in future iterations, that is, the probability of those weight adjustment actions that performed well in historical iteration rounds being selected in the future will increase. Based on the current iteration round and the first return values of each weight adjustment action, the prediction system can adopt an intelligent decision-making mechanism to select the target adjustment action. The above intelligent decision-making mechanism comprehensively considers the balance between exploration (trying new actions to discover potential optimization opportunities) and exploitation (selecting those actions that are known to improve the prediction performance), ensuring that the prediction system can both discover better weight configurations and further improve the accuracy of the prediction model on the existing basis. After the prediction system selects the target adjustment action, the above initial weights will be adjusted accordingly to obtain the new first adjusted weight and the second adjusted weight. The above adjustment process may include increasing or decreasing the weight values of the above two models, aiming to improve the fusion effect of the prediction results of the adjusted recurrent neural network and deep convolutional network models to more accurately reflect the actual power load characteristics. Using the first adjusted weight and the second adjusted weight, the prediction system can generate a new load prediction sequence, which is then matched and analyzed with the actual load data sequence to evaluate the matching degree between the prediction result and the real data. If the new load prediction sequence matches the load data sequence well, that is, the prediction error is significantly reduced, then the first adjusted weight and the second adjusted weight will be officially determined by the prediction system as the first model weight and the second model weight for subsequent prediction processes.

[0092] Further, obtaining the first return values of multiple weight adjustment actions in the weight adjustment action set at the current iteration round includes: obtaining the second return value of the target adjustment action selected in the previous iteration round and the load matching result of the previous iteration round; based on the second return value and the load matching result, adjusting the current return values of the multiple weight adjustment actions to obtain the first return value.

[0093] The above second return value may refer to the return value obtained by the target adjustment action selected by the prediction system in the previous iteration round. The above second return value reflects the influence of the first adjustment weight and the second adjustment weight obtained by adjusting the initial weight on the load prediction sequence in the previous iteration round.

[0094] In an alternative embodiment, the prediction system can obtain the second return value of the target adjustment action selected in the previous iteration round and the load matching result of the previous iteration round. The process of obtaining the above second return value is actually to establish a dynamic feedback mechanism. By evaluating the prediction effect after the weight adjustment action is executed in the previous iteration round, the performance change of the prediction model is quantified. The above second return value not only reflects the magnitude of the prediction error, but also takes into account the comprehensive evaluation of the load matching result, including multiple factors such as the compliance of the prediction trend, the accuracy of the predicted value, and the stability of the prediction sequence. The above dynamic feedback mechanism ensures that the prediction system can adaptively adjust the weight adjustment strategy according to the performance of the prediction model, and continuously improve the prediction performance. When the prediction system adjusts the current return value of the weight adjustment action based on the second return value and the load matching result, the IQL algorithm can help the prediction system intelligently identify which adjustment actions are more effective in improving the prediction accuracy. By increasing the selection probability of these effective actions, the prediction system can guide the recurrent neural network and the deep convolutional network models to have better adaptability in the face of changing load characteristics. Even when facing load mutations or being affected by complex external factors, the prediction system can maintain the stability and accuracy of the prediction through timely weight adjustment.

[0095] Further, the method further includes: obtaining multiple data sequences of different initial influencing factors and multiple load data sequences of the power system in multiple time periods, where the initial influencing factors include key influencing factors; determining multiple mutual information values of different initial influencing factors based on the multiple data sequences and the multiple load data sequences, where different mutual information values correspond to different data sequences and different load data sequences; performing normalization processing on the multiple mutual information values to obtain the maximum information coefficient corresponding to different initial influencing factors; and selecting key influencing factors from the initial influencing factors based on the maximum information coefficient, where the maximum information coefficient of the key influencing factors is greater than the maximum information coefficients of other influencing factors, and the other influencing factors are used to represent the influencing factors other than the key influencing factors among the initial influencing factors.

[0096] The above initial influencing factors can refer to a series of potential factors that may affect electricity demand in electricity load forecasting. For example, the above initial influencing factors can include at least one or more of the following: meteorology, time, economic activities, electricity prices, social behaviors, etc., but are not limited thereto. The above mutual information value can be a numerical value used to evaluate the mutual dependence relationship between the initial influencing factors and the electricity load data sequence. The above maximum information coefficient can be a coefficient used to detect and quantify the potential non-linear relationship between random variables.

[0097] In an alternative embodiment, by obtaining multiple data sequences of different initial influencing factors and multiple load data sequences of the power system over multiple time periods, the recurrent neural network and the deep convolutional network model can learn the patterns of load changes at different time scales, thereby better predicting future load demands. In addition, the initial influencing factors cover factors that may affect the electricity load, including meteorological conditions, time factors, economic activities, social behaviors, etc. This can ensure that the recurrent neural network and the deep convolutional network model can comprehensively consider the relevant factors that affect the electricity load, improving the comprehensiveness and accuracy of the prediction. Further considering that the mutual information value can quantify the degree of information sharing between the initial influencing factors and the load data sequence, revealing which factors have the most significant impact on the electricity load change. Therefore, the prediction system can identify the key influencing factors that contribute the most to the electricity load change by calculating the mutual information values between different initial influencing factors and the load data sequence. The above key factors have a high mutual information value, indicating a strong correlation between the above key factors and the load data sequence, which has a significant impact on the prediction result. The prediction system can normalize the mutual information value and convert it into a maximum information coefficient, enabling a standardized comparison of the importance between different initial influencing factors and avoiding evaluation biases caused by variable scale differences. This step provides a unified evaluation standard for the screening of key influencing factors. By selecting key influencing factors based on the maximum information coefficient, the prediction model will focus on those input variables that have a greater impact on load forecasting, thereby improving the accuracy and efficiency of the model.

[0098] For example, the prediction system can use mutual information and grid partitioning to calculate the maximum information coefficient. Specifically, it can be assumed that the meteorological variable is Ω = {a i}(i = 1, 2, …, n), and the load sequence is L = {l j}(j = 1, 2, …, m), where n and m are the number of weather factors and the number of load sequences respectively. The calculation method of the mutual trust value can be shown as the following formula:

[0099]

[0100] In the formula, f MI (Ω, L) is the mutual trust value, p(ai ,l j ) is the joint probability density, and p(a i ) and p(l j ) are the marginal probability densities. Subsequently, the prediction system can define a grid, denoted as G = (X, Y), which divides the values of a i and l j in the dataset into two grids, denoted as X and Y respectively, and calculate the maximum mutual information value of the dataset D under the grid G division as shown in the following formula:

[0101]

[0102] In the formula, is the maximum mutual information value of the dataset D under the grid G division, and f MI (D|G) is the mutual information value of D within G. For a fixed dataset D, normalize all the maximum mutual information values of different grids G on D to the interval (0, 1). The maximum information coefficient is the normalized maximum mutual information value, and the calculation method of the maximum information coefficient can be shown as follows:

[0103]

[0104] In the formula, f MIC (D) is the maximum information coefficient, and L(γ) is a function of γ.

[0105] For ease of understanding, Figure 4 is a flowchart of an optional load prediction process based on a deep learning hybrid model according to an embodiment of the present invention. As Figure 4 shown, first, perform initialization, that is, establish a state set S and an action set A. Use the weights of BiGRU as the environmental state, divide [0, 1] into 10001 states at intervals of 0.0001 as the state set S. Establish a Q table, and the action set A = [-0.0001, 0.0001], where the rows represent states and the columns represent actions. Subsequently, establish a loss function and a reward and punishment mechanism R. The establishment process can be shown as follows:

[0106]

[0107] In the formula, Y(t) represents the actual load value, represents the predicted load value, is the number of sample points, is the loss function value obtained by the m-th action operation.

[0108] Then, execute the action selection mechanism according to the foregoing formula and update the Q-table according to the foregoing formula. Then, determine whether the Q-table converges. If it converges and converges in the next 1000 learning iterations, output the weight υ1, and the weight of the deep residual convolution model is 1 - υ1. If it does not converge, repeat the execution of the action selection mechanism and update the Q-table until the Q-table converges. Finally, add up the prediction results of all load subsequences to obtain the final load prediction result.

[0109] According to an embodiment of the present invention, an embodiment of a power load prediction device is provided. It should be noted that this device can be used to execute the above-mentioned power load prediction method. The specific implementation manner and application scenario are the same as those of the above embodiment and will not be elaborated here. Figure 5 is a schematic diagram of a power load prediction device according to an embodiment of the present invention, as Figure 5 shown, the device includes:

[0110] A first acquisition module 502, configured to acquire the original load data sequence of the power system and the influence data sequence of key influencing factors within a historical time period, where the key influencing factors are used to characterize the factors that will affect the load data sequence of the power system.

[0111] A first decomposition module 504, configured to perform iterative decomposition on the original load data sequence to obtain at least one load sub-data sequence, where the load sub-data sequences corresponding to different iterative rounds are different.

[0112] A first prediction module 506, configured to input at least one load sub-data sequence and the influence data sequence into a load prediction model set, and use the load prediction model set to perform compliance prediction on the power system to obtain the target load data sequence of the power system within a future time period.

[0113] Further, the first decomposition module is further configured to: for any one of multiple iterative rounds, acquire the noise signal and the residual load data sequence corresponding to the iterative round, where the residual load data sequence is used to characterize the load data sequence in the original load data sequence except for the load sub-data sequence corresponding to the historical iterative round, and the residual load data sequence of the first iterative round is the original load data sequence, and the historical iterative round is the iterative round executed before the iterative round in multiple iterative rounds; add the noise signal to the residual load data sequence to obtain a noise load data sequence; perform signal decomposition on the noise load data sequence to obtain the load sub-data sequence corresponding to the iterative round.

[0114] Further, the first decomposition module is further configured to: obtain a noise component and a signal-to-noise ratio corresponding to an iteration round, and a standard deviation of the residual load data sequence; obtain a noise adjustment coefficient based on a ratio of the signal-to-noise ratio to the standard deviation; and adjust the noise component based on the noise adjustment coefficient to obtain a noise signal.

[0115] Further, the load prediction model set includes: a recurrent neural network model and a deep convolutional network model; the first prediction module is further configured to: input at least one load sub-data sequence and an influence data sequence into the recurrent neural network model to obtain a first load data sequence; input at least one load sub-data sequence and the influence data sequence into the deep convolutional network model to obtain a second load data sequence; and perform a weighted process on the first load data sequence and the second load data sequence based on a first model weight corresponding to the recurrent neural network model and a second model weight corresponding to the deep convolutional network model to obtain a target load data sequence.

[0116] Further, the apparatus further includes: a second acquisition module, configured to acquire a training data sequence in a first time period and a load data sequence in a second time period, where an end time of the first time period is before a start time of the second time period, the training data sequence includes: a load data sequence of the power system in the first time period and a data sequence of key influencing factors in the first time period, and the load data sequence includes: a load data sequence of the power system in the first time period; a second prediction module, configured to predict the load data sequence of the power system in the second time period by using the recurrent neural network model based on the training data sequence to obtain a first prediction sequence, and predict the load data sequence of the power system in the second time period by using the deep convolutional network model based on the training data sequence to obtain a second prediction sequence; a first calculation module, configured to perform a weighted process on the first prediction sequence and the second prediction sequence based on initial weights to obtain a load prediction sequence, where the initial weights include: a first initial weight corresponding to the recurrent neural network model and a second initial weight corresponding to the deep convolutional network model; a first matching module, configured to match the load prediction sequence with the load data sequence to obtain a sequence matching result; and a first adjustment module, configured to adjust the initial weights based on the sequence matching result to obtain a first model weight and a second model weight.

[0117] Further, the first adjustment module is further configured to: in response to the sequence matching result indicating that the load prediction sequence does not match the load data sequence, obtain the current iteration round and the first return values of multiple weight adjustment actions in the weight adjustment action set in the current iteration round, where the first return values are used to represent the selection probabilities of different weight adjustment actions; based on the current iteration round and the first return values, select a target adjustment action from the weight adjustment action set; adjust the initial weights based on the target adjustment action to obtain a first adjusted weight and a second adjusted weight; based on the first adjusted weight and the second adjusted weight, re-determine the load prediction sequence, and when the new load prediction sequence matches the load data sequence, determine the first adjusted weight as the first model weight and determine the second adjusted weight as the second model weight.

[0118] Further, obtaining the first return values of multiple weight adjustment actions in the weight adjustment action set in the current iteration round includes: obtaining the second return value of the target adjustment action selected in the previous iteration round and the load matching result in the previous iteration round; adjusting the current return values of the multiple weight adjustment actions based on the second return value and the load matching result to obtain the first return values.

[0119] Further, the method further includes: obtaining multiple data sequences of different initial influencing factors and multiple load data sequences of the power system in multiple time periods, where the initial influencing factors include key influencing factors; determining multiple mutual information values of different initial influencing factors based on the multiple data sequences and the multiple load data sequences, where different mutual information values correspond to different data sequences and different load data sequences; performing normalization processing on the multiple mutual information values to obtain the maximum information coefficients corresponding to different initial influencing factors; selecting key influencing factors from the initial influencing factors based on the maximum information coefficients, where the maximum information coefficient of the key influencing factors is greater than the maximum information coefficients of other influencing factors, and the other influencing factors are used to represent the influencing factors other than the key influencing factors in the initial influencing factors.

[0120] An embodiment of the present application further provides an electronic device, including: a memory storing an executable program; a processor configured to run the program, where when the program runs, it executes the methods in various embodiments of the present invention.

[0121] An embodiment of the present application further provides a computer-readable storage medium, where the computer-readable storage medium includes a stored executable program, and when the executable program runs, it controls the device where the computer-readable storage medium is located to execute the methods in various embodiments of the present invention.

[0122] An embodiment of the present application further provides a computer program product, including a computer program that implements the methods in various embodiments of the present invention when executed by a processor.

[0123] An embodiment of the present application further provides a computer program product, including a non-volatile computer-readable storage medium for storing a computer program, where the computer program, when executed by a processor, implements the methods in various embodiments of the present invention.

[0124] An embodiment of the present application further provides a computer program, where the computer program, when executed by a processor, implements the methods in various embodiments of the present invention described above.

[0125] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0126] In the above embodiments of the present invention, the descriptions of the various embodiments have their own emphases. For the parts not detailed in a certain embodiment, reference may be made to the relevant descriptions of other embodiments.

[0127] In several embodiments provided by the present application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only illustrative. For example, the division of the units can be a logical function division. In actual implementation, there can be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling or direct coupling or communication connection between each other can be through some interfaces. The indirect coupling or communication connection of units or modules can be in an electrical or other form.

[0128] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they can be located in one place or distributed to multiple units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0129] In addition, the functional units in various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above integrated units can be implemented in the form of hardware or in the form of software functional units.

[0130] When the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, read-only memories (ROMs), random access memories (RAMs), mobile hard disks, magnetic disks, or optical discs that can store program codes.

[0131] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A method for predicting power load, characterized in that: include: Acquire the original load data sequence of the power system and the impact data sequence of key influencing factors within a historical time period, wherein the key influencing factors are used to characterize the factors that may affect the load data sequence of the power system; Iteratively decomposing the original load data sequence to obtain at least one load sub-data sequence, wherein different load sub-data sequences corresponding to different iteration rounds are different; The at least one load sub-data sequence and the impact data sequence are input into a load forecasting model set, and the load forecasting model set is used to perform compliance forecasting on the power system to obtain a target load data sequence of the power system in a future time period.

2. The method according to claim 1, characterized in that The original load data sequence is iteratively decomposed to obtain at least one load sub-data sequence, including: For any one of the multiple iteration rounds, a noise signal and a residual load data sequence corresponding to the iteration round are obtained, wherein the residual load data sequence is used to characterize the load data sequence in the original load data sequence except the load sub-data sequence corresponding to the historical iteration round, the residual load data sequence of the first iteration round is the original load data sequence, and the historical iteration round is the iteration round executed before the iteration round in the multiple iteration rounds; Adding the noise signal to the residual load data sequence to obtain a noise load data sequence; Signal decomposition is performed on the noise load data sequence to obtain a load sub-data sequence corresponding to the iteration round.

3. The method according to claim 2, characterized in that Acquiring a noise signal corresponding to the iteration round includes: Obtaining the noise component and signal-to-noise ratio corresponding to the iteration round, and the standard deviation of the residual load data sequence; Obtaining a noise adjustment factor based on a ratio of the signal-to-noise ratio to the standard deviation; The noise component is adjusted based on the noise adjustment coefficient to obtain the noise signal.

4. The method according to claim 1, characterized in that: The load forecasting model set includes: a recurrent neural network model and a deep convolutional network model; the at least one load sub-data sequence and the impact data sequence are input into the load forecasting model set, and the load forecasting model set is used to perform compliance forecasting on the power system to obtain a target load data sequence of the power system in a future time period, including: Inputting the at least one load sub-data sequence and the influence data sequence into the recurrent neural network model to obtain a first load data sequence; Inputting the at least one load sub-data sequence and the impact data sequence into the deep convolutional network model to obtain a second load data sequence; Based on the first model weight corresponding to the recurrent neural network model and the second model weight corresponding to the deep convolutional network model, the first load data sequence and the second load data sequence are weighted to obtain the target load data sequence.

5. The method according to claim 4, characterized in that The method further comprises: Acquire a training data sequence of a first time period and a load data sequence of a second time period, wherein the end time of the first time period is before the start time of the second time period, the training data sequence includes: a load data sequence of the power system in the first time period and a data sequence of the key influencing factors in the first time period, and the load data sequence includes: a load data sequence of the power system in the first time period; Based on the training data sequence, using a recurrent neural network model to predict the load data sequence of the power system in the second time period to obtain a first prediction sequence, and based on the training data sequence, using the deep convolutional network model to predict the load data sequence of the power system in the second time period to obtain a second prediction sequence; Performing weighted processing on the first prediction sequence and the second prediction sequence based on initial weights to obtain a load prediction sequence, wherein the initial weights include: a first initial weight corresponding to the recurrent neural network model and a second initial weight corresponding to the deep convolutional network model; Matching the load forecast sequence with the load data sequence to obtain a sequence matching result; The initial weight is adjusted based on the sequence matching result to obtain the first model weight and the second model weight.

6. The method according to claim 5, characterized in that Adjusting the first initial weight and the second initial weight based on the sequence matching result to obtain the first model weight and the second model weight includes: In response to the sequence matching result being that the load forecast sequence does not match the load data sequence, obtaining a current iteration round and first reward values ​​of a plurality of weight adjustment actions in a weight adjustment action set in the current iteration round, wherein the first reward value is used to characterize the selection probability of different weight adjustment actions; Based on the current iteration round and the first reward value, selecting a target adjustment action from a weight adjustment action set; Adjusting the initial weight based on the target adjustment action to obtain a first adjustment weight and a second adjustment weight; Based on the first adjustment weight and the second adjustment weight, the load forecast sequence is redetermined, and when the new load forecast sequence matches the load data sequence, the first adjustment weight is determined to be the first model weight, and the second adjustment weight is determined to be the second model weight.

7. The method according to claim 6, characterized in that Get the first reward value of multiple weight adjustment actions in the weight adjustment action set in the current iteration, including: Obtaining a second reward value of the target adjustment action selected in the previous iteration round and a load matching result of the previous iteration round; Based on the second reward value and the load matching result, current reward values ​​of the multiple weight adjustment actions are adjusted to obtain the first reward value.

8. The method according to claim 1, characterized in that The method further comprises: Acquire multiple data sequences of different initial influencing factors in multiple time periods, and multiple load data sequences of the power system, wherein the initial influencing factors include the key influencing factors; Determining a plurality of mutual information values ​​of different initial influencing factors based on the plurality of data sequences and the plurality of load data sequences, wherein different mutual information values ​​correspond to different data sequences and different load data sequences; Normalizing the multiple mutual information values ​​to obtain maximum information coefficients corresponding to different initial influencing factors; Based on the maximum information coefficient, the key influencing factor is selected from the initial influencing factors, wherein the maximum information coefficient of the key influencing factor is greater than the maximum information coefficients of other influencing factors, and the other influencing factors are used to characterize the influencing factors in the initial influencing factors except the key influencing factor.

9. An electronic device, characterized in that: include: A memory storing an executable program; A processor, configured to run the program, wherein the program executes the method according to any one of claims 1 to 6 when running.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored executable program, wherein when the executable program is executed, the device where the storage medium is located is controlled to execute the method according to any one of claims 1 to 6.

Images