Electric power resource prediction method and device, computer equipment and storage medium

Abstract

The invention relates to an electric power resource prediction method and device, computer equipment and a storage medium, and relates to the technical field of electric power. The method comprises the following steps: decomposing a power resource sequence based on a wavelet transform strategy to obtain a stationary sequence and a non-stationary sequence; predicting the stationary sequence based on a preset random time sequence model to obtain a first prediction result; the random time sequence model comprises an AR sub-model and an MA sub-model; the AR sub-model and the MA sub-model are determined based on an autocorrelation function and a partial autocorrelation function corresponding to a historical power resource sequence; predicting the non-stationary sequence based on a preset kernel extreme learning machine KELM model to obtain a second prediction result; the KELM model is obtained by training based on a historical power resource sequence and a simulated annealing particle swarm optimization SAPSO strategy; and obtaining a target prediction result corresponding to the power resource sequence based on the first prediction result and the second prediction result. The method can improve the prediction accuracy of electric power resources.

Description

Technical Field

[0001] The present application relates to the field of power technology, and particularly to a power resource prediction method, device, computer device, storage medium, and computer program product. Background Art

[0002] Power resources in the power market are affected by various factors such as supply and demand relationships, climatic conditions, and fuel prices, and have strong volatility and uncertainty. Power resource prediction is the basis for rational decision-making by various participants in the power market, and plays a key role in reducing market risks, improving energy efficiency, and supporting the development of renewable energy. In the power spot market, power generation and selling enterprises on the supply side can formulate power clearing or purchasing strategies based on the predicted power resources, and power-consuming enterprises on the demand side can optimize production plans by predicting the trend of power resources, thereby reducing electricity costs.

[0003] Currently, various time series models are used in related technologies for power resource prediction, and the prediction ability of the model is determined by fitting the power resource data of the power market at a specific spatio-temporal scale through the time series model. The time series model has good performance in processing time series data and can capture the periodic, trend, and seasonal characteristics of power resources. However, it has limited performance for non-linear and complex fluctuating power resource data and cannot fully capture the non-linear characteristics of power resources affected by external factors (such as weather, fuel prices, etc.), resulting in insufficient prediction accuracy. Summary of the Invention

[0004] Based on this, it is necessary to provide a power resource prediction method, device, computer device, computer-readable storage medium, and computer program product that can improve prediction accuracy for the above technical problems.

[0005] In a first aspect, the present application provides a power resource prediction method. The method includes:

[0006] Decompose the power resource sequence based on a wavelet transform strategy to obtain a stationary sequence and a non-stationary sequence;

[0007] Predict the stationary sequence based on a preset stochastic time series model to obtain a first prediction result; the stochastic time series model includes an AR sub-model and an MA sub-model; the AR sub-model and the MA sub-model are determined based on the autocorrelation function and partial autocorrelation function corresponding to the historical power resource sequence;

[0008] Predict the non-stationary sequence based on a preset kernel extreme learning machine KELM model to obtain a second prediction result; the KELM model is trained based on the historical power resource sequence and a simulated annealing particle swarm optimization SAPSO strategy;

[0009] Based on the first prediction result and the second prediction result, obtain the target prediction result corresponding to the power resource sequence.

[0010] In one embodiment, the wavelet transform strategy includes a multi-resolution analysis fast algorithm; decomposing the power resource sequence based on the wavelet transform strategy to obtain a stationary sequence and a non-stationary sequence, including:

[0011] Filter the power resource sequence based on the low-pass filter in the multi-resolution analysis fast algorithm to obtain an approximation sequence; the approximation sequence represents the trend information of the power resource sequence;

[0012] Filter the power resource sequence based on the high-pass filter in the multi-resolution analysis fast algorithm to obtain a detail sequence; the detail sequence represents the characteristic of influencing factors related to the power resource sequence;

[0013] Based on the autocorrelation function and the unit root test strategy of the power resource sequence, perform stationarity analysis on the approximation sequence and the detail sequence respectively to determine the sequence types of the approximation sequence and the detail sequence; the sequence types include stationary sequences and non-stationary sequences.

[0014] In one embodiment, predicting the stationary sequence based on a preset random time series model to obtain a first prediction result, including:

[0015] According to the time of each element in the stationary sequence, input each element into the AR sub-model and the MA sub-model in the random time series model to obtain the prediction result of the AR sub-model and the prediction result of the MA sub-model;

[0016] Add the prediction result of the AR sub-model and the prediction result of the MA sub-model to obtain the first prediction result corresponding to the stationary sequence.

[0017] In one embodiment, predicting the non-stationary sequence based on a preset kernel extreme learning machine KELM model to obtain a second prediction result, including:

[0018] Input the non-stationary sequence into the kernel mapping layer of the KELM model to obtain the kernel vector corresponding to the non-stationary sequence;

[0019] Perform weighted summation on the output weight of the KELM model and the kernel vector to obtain the second prediction result corresponding to the non-stationary sequence.

[0020] In one embodiment, the method further includes:

[0021] Obtain the historical stationary sequence in the historical power resource sequence;

[0022] Determine the ARMA model based on the autocorrelation function and partial autocorrelation function corresponding to the historical stationary sequence;

[0023] Estimate each parameter in the ARMA model to obtain an updated ARMA model, and determine the white noise test result of the updated ARMA model;

[0024] If the residual sequence in the white noise test result is not a white noise sequence, return to execute the step of estimating each parameter in the ARMA model to obtain an updated ARMA model until the residual sequence in the white noise test result is a white noise sequence, then determine the updated ARMA model as the random time series model.

[0025] In one embodiment, the method further includes:

[0026] Determine the particles in the SAPSO strategy as the regularization parameter and kernel parameter of the KELM model;

[0027] Determine the fitness value corresponding to the regularization parameter and kernel parameter based on the fitness function;

[0028] Update the local best position, global best position, velocity and position of each particle according to the fitness value until the training termination condition is met, and obtain the optimal regularization parameter and kernel parameter;

[0029] Determine the trained KELM model based on the optimal regularization parameter and kernel parameter.

[0030] In one embodiment, the method further includes:

[0031] Based on a preset performance evaluation strategy, evaluate the target prediction result and the actual result corresponding to the target prediction result to obtain the prediction performance of the random time series model and the KELM model; the performance evaluation strategy includes at least one of mean absolute error, mean absolute percentage error and root mean square error.

[0032] In a second aspect, the present application also provides a power resource prediction device. The device includes:

[0033] A sequence decomposition module, configured to decompose the power resource sequence based on a wavelet transform strategy to obtain a stationary sequence and a non-stationary sequence;

[0034] The first prediction module is used to predict the stationary sequence based on a preset random time series model to obtain a first prediction result; the random time series model includes an AR sub-model and an MA sub-model; the AR sub-model and the MA sub-model are determined based on the autocorrelation function and partial autocorrelation function corresponding to the historical power resource sequence;

[0035] The second prediction module is used to predict the non-stationary sequence based on a preset Kernel Extreme Learning Machine (KELM) model to obtain a second prediction result; the KELM model is trained based on the historical power resource sequence and the Simulated Annealing Particle Swarm Optimization (SAPSO) strategy;

[0036] The result determination module is used to obtain the target prediction result corresponding to the power resource sequence based on the first prediction result and the second prediction result.

[0037] In a third aspect, the present application also provides a computer device. The computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the method described in the first aspect are implemented.

[0038] In a fourth aspect, the present application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method described in the first aspect are implemented.

[0039] In a fifth aspect, the present application also provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the steps of the method described in the first aspect are implemented.

[0040] The above power resource prediction method, device, computer device, storage medium, and computer program product decompose a power resource sequence based on a wavelet transform strategy to obtain a stationary sequence and a non-stationary sequence, predict the stationary sequence based on a preset stochastic time series model to obtain a first prediction result, where the stochastic time series model includes an AR sub-model and an MA sub-model, and the AR sub-model and the MA sub-model are determined based on the autocorrelation function and partial autocorrelation function corresponding to the historical power resource sequence. Then, predict the non-stationary sequence based on a preset Kernel Extreme Learning Machine (KELM) model to obtain a second prediction result, and the KELM model is trained based on the historical power resource sequence and a Simulated Annealing Particle Swarm Optimization (SAPSO) strategy. Finally, based on the first prediction result and the second prediction result, obtain the target prediction result corresponding to the power resource sequence, which can decompose the power resource sequence into a stationary sequence and a non-stationary sequence, learn the features in the stationary sequence through the stochastic time series model, and predict the first prediction result of the stationary sequence in the future. At the same time, learn the features in the non-stationary sequence through the KELM model and predict the second prediction result of the non-stationary sequence in the future, which can fully capture the periodic, trend, and seasonal features in the stationary sequence, as well as capture the non-linear features in the non-stationary sequence, thereby improving the accuracy of the target prediction result. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] To more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following will briefly introduce the drawings required for use in the description of the embodiments or related technologies. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0042] Figure 1 It is an application environment diagram of the power resource prediction method in an embodiment;

[0043] Figure 2 It is a flowchart of the power resource prediction method in an embodiment;

[0044] Figure 3 It is a flowchart of the steps for obtaining a stationary sequence and a non-stationary sequence in an embodiment;

[0045] Figure 4 It is a flowchart of the power resource prediction method in another embodiment;

[0046] Figure 5 It is a flowchart of the Mallat algorithm for decomposing a power resource signal in an embodiment;

[0047] Figure 6It is a structural block diagram of a power resource prediction device in an embodiment;

[0048] Figure 7 It is an internal structure diagram of a computer device in an embodiment. Specific implementation manners

[0049] In order to make the objectives, technical solutions and advantages of the present application clearer and more understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0050] The power resource prediction method provided by the embodiments of the present application can be applied to an application environment as Figure 1 shown. Among them, the terminal 102 communicates with the server 104 through a network. The terminal 102 can record the data of the power resources over a period of time and determine it as a power resource sequence. Each terminal 102 can send its own power resource sequence to the server 104, and the server 104 can integrate the power resource sequences. The server 104 can classify and process the power resource sequences according to requirements to obtain a power resource sequence that meets the user's needs. The server 104 can determine the power resource sequence within a preset time period as the historical power resource sequence according to the requirements of the training model. In addition, the server can decompose the power resource sequence to be predicted to obtain multiple subsequences and determine whether each subsequence is a stationary sequence or a non-stationary sequence. The server 104 can predict the stationary sequence based on a preset random time series model to obtain a first prediction result of the stationary sequence at a future time. The server 104 can predict the non-stationary sequence based on a preset Kernel Extreme Learning Machine (KELM) model to obtain a second prediction result of the non-stationary sequence at a future time. The server 104 combines the first prediction result and the second prediction result to obtain a target prediction result of the power resource sequence at a future time. Among them, the random time series model includes an Auto-Regressive (AR) sub-model and a Moving Average (MA) sub-model. The AR sub-model is an autoregressive sub-model, and the MA sub-model is a moving average sub-model. The KELM model is trained based on the historical power resource sequence and the Simulated Annealing Particle Swarm Optimization (SAPSO) strategy.

[0051] The data storage system can store the data that server 104 needs to process. The data storage system can be integrated on server 104, or placed on the cloud or other network servers. Among them, terminal 102 can be, but is not limited to, various personal computers, laptops, power devices, and Internet of Things devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. Server 104 can be implemented by an independent server or a server cluster composed of multiple servers.

[0052] In an exemplary embodiment, as Figure 2 shown, a power resource prediction method is provided. Taking the method applied to Figure 1 server 104 in it as an example, the following steps S202 to step S208 are included. Among them:

[0053] Step S202, decompose the power resource sequence based on the wavelet transform strategy to obtain a stationary sequence and a non-stationary sequence.

[0054] Among them, wavelet transform is a mathematical tool for signal analysis, which can decompose a signal into different frequency components and analyze the performance of each component at different time points. The wavelet transform strategy can include two types, such as Continuous Wavelet Transform (CWT) and Discrete Wavelet Transform (DWT). A stationary sequence can be that any finite-dimensional distribution of a time series does not change over time, and a non-stationary sequence can be a time series that does not meet the definition of a stationary sequence, that is, its statistical characteristics change over time. For example, the mean, variance, and autocovariance of a non-stationary sequence can change over time. The power resource can be the amount of electricity generated within a certain period of time, the amount of electricity consumed within a certain period of time, or the price corresponding to the amount of electricity generated or consumed within a certain period of time. This is not specifically limited in the embodiments of the present application. The power resource usually changes continuously over time. For example, there are fluctuations in the power resources during peak electricity consumption periods, peak electricity consumption seasons, off-peak electricity consumption periods, and off-peak electricity consumption seasons.

[0055] Specifically, the server can decompose the power resource sequence to be predicted according to a preset wavelet change strategy, such as the Mallat algorithm, to obtain multiple subsequences, and determine whether each subsequence is a stationary sequence or a non-stationary sequence. For the subsequences that are all stationary sequences, the server can merge them to obtain a stationary sequence composed of multiple subsequences. For the subsequences that are all non-stationary sequences, the server can merge them to obtain a non-stationary sequence composed of multiple subsequences.

[0056] Step S204: Predict the stationary sequence based on a preset random time series model to obtain a first prediction result.

[0057] Among them, the random time series model can be an Auto-Regressive and Moving Average (ARMA) model. The random time series model includes an AR sub-model and an MA sub-model. The AR sub-model and the MA sub-model are determined based on the autocorrelation function and the partial autocorrelation function corresponding to the historical power resource sequence. The autocorrelation function measures the linear correlation between the observed values at different time points in the time series. In the ARMA model, the autocorrelation function can help determine the characteristics of the time series. The partial autocorrelation function measures the linear correlation between two observed values in the time series after eliminating the influence of the intermediate lag terms.

[0058] Specifically, the server can determine the autocorrelation function and the partial autocorrelation function corresponding to the historical power resource sequence, and determine the p value and the q value in the random time series model according to the truncation and trailing characteristics corresponding to the autocorrelation function and the partial autocorrelation function respectively, and determine the AR sub-model and the MA sub-model according to the p value and the q value. After constructing the ARMA model, the server can input the stationary sequence corresponding to the power resource sequence to be predicted into the ARMA model. The ARMA model can output the first prediction result according to the constructed AR sub-model and MA sub-model, and obtain the predicted value of the power resource in the future time.

[0059] Step S206: Predict the non-stationary sequence based on a preset Kernel Extreme Learning Machine (KELM) model to obtain a second prediction result.

[0060] Among them, the KELM model is trained based on the historical power resource sequence and the Simulated Annealing Particle Swarm Optimization (SAPSO) strategy.

[0061] Specifically, the server can train the KELM model according to the historical power resource sequence and the SAPSO strategy. After training the KELM model, input the non-stationary sequence corresponding to the power resource sequence to be predicted into the KELM model to obtain a second prediction result, so as to obtain the predicted value of the power resource in the future time.

[0062] Step S208: Obtain the target prediction result corresponding to the power resource sequence based on the first prediction result and the second prediction result.

[0063] Specifically, the server can perform corresponding weighted summation on the first prediction result and the second prediction result in chronological order. The server can obtain the predicted power resource values at the same time point in the first prediction result and the predicted power resource values at the same time point in the second prediction result, and perform weighted summation on the predicted power resource values in different prediction results according to the preset weight values to obtain the target predicted value corresponding to each future time. Determining the target predicted value for a period of time in the future as the target prediction result corresponding to the power resource sequence to be predicted.

[0064] In the above power resource prediction method, by decomposing the power resource sequence based on the wavelet transform strategy, a stationary sequence and a non-stationary sequence are obtained. Based on the preset random time series model, the stationary sequence is predicted to obtain the first prediction result. Among them, the random time series model includes an AR sub-model and an MA sub-model, and the AR sub-model and the MA sub-model are determined based on the autocorrelation function and the partial autocorrelation function corresponding to the historical power resource sequence. Then, based on the preset kernel extreme learning machine KELM model, the non-stationary sequence is predicted to obtain the second prediction result. The KELM model is trained based on the historical power resource sequence and the simulated annealing particle swarm optimization SAPSO strategy. Finally, based on the first prediction result and the second prediction result, the target prediction result corresponding to the power resource sequence is obtained, which can decompose the power resource sequence into a stationary sequence and a non-stationary sequence, learn the characteristics in the stationary sequence through the random time series model, and predict the first prediction result of the stationary sequence in the future. At the same time, learn the characteristics in the non-stationary sequence through the KELM model, and predict the second prediction result of the non-stationary sequence in the future, which can fully capture the periodic, trend and seasonal characteristics in the stationary sequence, as well as capture the non-linear characteristics in the non-stationary sequence, thereby improving the accuracy of the target prediction result.

[0065] In an exemplary embodiment, the wavelet transform strategy includes a multi-resolution analysis fast algorithm, such as Figure 3 shown, the specific implementation process of the step "decompose the power resource sequence based on the wavelet transform strategy to obtain a stationary sequence and a non-stationary sequence" includes steps S302 to S306. Among them:

[0066] Step S302, filter the power resource sequence based on the low-pass filter in the multi-resolution analysis fast algorithm to obtain an approximation sequence.

[0067] Specifically, the wavelet transform strategy includes fast algorithms for multi-resolution analysis, usually referring to the fast algorithms in wavelet analysis, such as the Mallat algorithm. The approximation sequence represents the trend information of the power resource sequence. When the time series passes through a low-pass filter, a new sequence is obtained. This sequence retains the low-frequency components in the original time series, reflecting the overall trend and slow-changing characteristics of the sequence, and this sequence is the approximation sequence. For example, for the time series of power resource trading volume, the approximation sequence can reflect the long-term trend of the power resource trading volume. The server filters the power resource sequence to be predicted through the low-pass filter in the Mallat algorithm, retaining the low-frequency components, so as to obtain the approximation sequence corresponding to the power resource sequence.

[0068] Step S304: Filter the power resource sequence based on the high-pass filter in the fast algorithm for multi-resolution analysis to obtain the detail sequence.

[0069] Specifically, the detail sequence represents the characteristic of influencing factors related to the power resource sequence. After the time series is processed by the high-pass filter, a sequence containing high-frequency components is obtained. These high-frequency components represent the fast changes and local detail information in the sequence, and this sequence is the detail sequence. In the example of power resource trading volume, the detail sequence can reflect the short-term fluctuations and sudden changes in the power resource trading volume. The server filters the power resource sequence to be predicted through the high-pass filter in the Mallat algorithm, retaining the high-frequency components, so as to obtain the detail sequence corresponding to the power resource sequence.

[0070] In one example, the server can use Daubechies5 in the Mallat algorithm as the mother wavelet to perform five-level decomposition on the signal corresponding to the power resource sequence, obtaining five sub-sequences. For example, if F(t) is the power resource sequence, after the first decomposition, d1(t) and c1(t) are obtained. The third-level decomposition is performed on c1(t) to obtain d2(t) and c2(t). The fourth-level decomposition is performed on c2(t) to obtain d3(t) and c3(t). The fifth-level decomposition is performed on c3(t) to obtain d4(t) and c4(t). At this time, five sub-sequences, d1(t), c1(t), c2(t), c3(t), and c4(t), can be obtained. The sequence corresponding to d1(t) is the approximation sequence, and the sequences corresponding to c1(t), c2(t), c3(t), and c4(t) are the detail sequences.

[0071] Step S306: Based on the autocorrelation function and unit root test strategy of the power resource sequence, perform stationarity analysis on the approximation sequence and the detail sequence respectively to determine the sequence types of the approximation sequence and the detail sequence.

[0072] Among them, the sequence types include stationary sequences and non-stationary sequences. The unit root test is a statistical test method to determine whether there is a unit root in the autocorrelation function of a time series. If there is a unit root, the time series is non-stationary; conversely, if there is no unit root, the time series is stationary.

[0073] Specifically, the server can determine the autocorrelation function of the approximate sequence and the autocorrelation function of the detail sequence, and determine whether the approximate sequence or the detail sequence is stationary by determining the decay rate of the autocorrelation function, and determine the sequence type of the approximate sequence or the detail sequence as a stationary sequence or a non-stationary sequence. The server can use the unit root test strategy to determine whether there is a unit root in each autocorrelation function. If there is a unit root, it is determined that the sequence corresponding to the autocorrelation function is non-stationary. If there is no unit root, it is determined that the sequence corresponding to the autocorrelation function is stationary.

[0074] In this embodiment, the power resource sequence is decomposed by the fast algorithm of multi-resolution analysis to obtain multiple approximate sequences and detail sequences, and the types of each sequence are determined respectively through the autocorrelation function and the unit root test strategy, and it is determined which sequences are stationary sequences and which sequences are non-stationary sequences, which can fully explore the internal characteristics of each sequence and improve the accuracy of the determined stationary sequences and non-stationary sequences.

[0075] In an exemplary embodiment, the specific implementation process of the step "predicting the stationary sequence based on a preset random time series model to obtain a first prediction result" includes:

[0076] According to the time of each element in the stationary sequence, each element is input into the AR sub-model and the MA sub-model in the random time series model to obtain the prediction result of the AR sub-model and the prediction result of the MA sub-model. The prediction result of the AR sub-model and the prediction result of the MA sub-model are added together to obtain the first prediction result corresponding to the stationary sequence.

[0077] Specifically, for each stationary sequence, the server can input each element into the AR sub-model and the MA sub-model in the random time series model according to the time sequence of each element in the stationary sequence. The AR sub-model can generate a prediction result corresponding to the AR sub-model based on the input elements. The MA sub-model can generate a prediction result corresponding to the MA sub-model based on the input elements. The server adds the prediction result corresponding to the AR sub-model and the prediction result corresponding to the MA sub-model to obtain the first prediction result corresponding to the stationary sequence.

[0078] In this embodiment, by combining the prediction results of the AR sub-model and the MA sub-model respectively, the first prediction result can be obtained, which can improve the prediction accuracy of the stationary sequence.

[0079] In an exemplary embodiment, the specific implementation process of the step "predicting the non-stationary sequence based on a preset Kernel Extreme Learning Machine (KELM) model to obtain a second prediction result" includes:

[0080] Input the non-stationary sequence into the kernel mapping layer of the KELM model to obtain the kernel vector corresponding to the non-stationary sequence; perform weighted summation on the output weights and the kernel vector of the KELM model to obtain the second prediction result corresponding to the non-stationary sequence.

[0081] Specifically, for each non-stationary sequence, the server can input each element into the input layer of the KELM model in the chronological order of the elements in the non-stationary sequence. The input layer transmits the data to the kernel mapping layer, and through the kernel function, the input data is mapped from the original space to a high-dimensional feature space to obtain the kernel vector of the non-stationary sequence in the high-dimensional feature space. The server can perform weighted summation on the kernel vector according to the output weights of the trained KELM model to obtain the second prediction result corresponding to the non-stationary sequence.

[0082] In an example, the kernel function in the KELM model can be a polynomial kernel, a hyperbolic kernel, or a Gaussian kernel.

[0083] In this embodiment, obtaining the second prediction result of the non-stationary sequence through the trained KELM model can improve the prediction accuracy of the non-stationary sequence.

[0084] In an exemplary embodiment, the method further includes:

[0085] Obtain the historical stationary sequence in the historical power resource sequence; determine an ARMA model based on the autocorrelation function and the partial autocorrelation function corresponding to the historical stationary sequence; estimate the parameters in the ARMA model to obtain an updated ARMA model, and determine the white noise test result of the updated ARMA model. If the residual sequence in the white noise test result is not a white noise sequence, return to execute the step of estimating the parameters in the ARMA model to obtain an updated ARMA model until the residual sequence in the white noise test result is a white noise sequence, then determine the updated ARMA model as the random time series model.

[0086] Specifically, the server can decompose the historical power resource sequence through the above-mentioned wavelet transform strategy to obtain the historical stationary sequence. Then, the server can determine the autocorrelation function and the partial autocorrelation function corresponding to the historical stationary sequence, initialize the parameters p and q according to the autocorrelation function and the partial autocorrelation function, and based on the values of the parameters p and q, establish an AR sub-model and an MA sub-model corresponding to the historical power resource sequence to obtain an initial ARMA model.

[0087] In an example, an ARMA(p, q) model with parameter orders p and q is defined as:

[0088]

[0089] Wherein, is an ARMA sequence, and the random term is an independent white noise sequence. is the autocorrelation coefficient, and θq is the moving average coefficient. If p = 0, the equation becomes a q-order MA sub-model; if q = 0, it becomes a p-order AR sub-model.

[0090] After that, the server can estimate the parameters in the ARMA model to obtain the updated parameters, thereby determining the updated ARMA model. For example, for the AR sub-model, the least squares method can be used to estimate the autoregressive coefficients. For the ARMA model, the extended least squares method or the non-linear least squares method can also be used for parameter estimation. Or, by constructing the likelihood function of the ARMA model and then maximizing the likelihood function to estimate the parameters of the model.

[0091] After determining the order and parameters of the ARMA model, it is necessary to test the model to ensure its effectiveness. Commonly used test methods include the white noise test of the residuals (such as the Ljung-Box test). If the residual sequence is a white noise sequence, it means that the model can fit the data well. It is also possible to perform a significance test of the parameters to determine whether the estimated parameters are significantly non-zero. If the model test fails, it is necessary to re-adjust the order of the model or use other methods for parameter estimation.

[0092] That is to say, the server can determine the white noise test result of the updated ARMA model. If the residual sequence in the white noise test result is not a white noise sequence, it returns to execute the steps of determining the parameters in the ARMA model to obtain the updated ARMA model, thereby repeating the parameter update until the residual sequence in the white noise test result is a white noise sequence, and then determining that the updated ARMA model is a random time series model.

[0093] In this embodiment, by establishing an ARMA model based on the historical stationary sequence in the historical power resource sequence, estimating the parameters of the ARMA model, and performing a white noise test, the prediction accuracy of the random time series model can be improved.

[0094] In an exemplary embodiment, the method further includes:

[0095] Determine the particles in the SAPSO strategy as the regularization parameter and kernel parameter of the KELM model; determine the fitness values corresponding to the regularization parameter and kernel parameter based on the fitness function; update the local best position, global best position, the velocity and position of each particle according to the fitness values until the training termination condition is met, and obtain the optimal regularization parameter and kernel parameter; determine the trained KELM model based on the optimal regularization parameter and kernel parameter.

[0096] Among them, the SAPSO strategy usually refers to the Simulated Annealing Particle Swarm Optimization algorithm, which is an optimization algorithm that combines the Simulated Annealing (SA) algorithm and the Particle Swarm Optimization (PSO) algorithm. The SAPSO algorithm uses the simulated annealing mechanism to improve the particle swarm algorithm: introducing the probability acceptance criterion of simulated annealing on the basis of the PSO algorithm. When the fitness value obtained after the particle updates its position is not as good as the previous position, instead of directly discarding this new position, it is accepted with a certain probability. This probability is related to the current temperature and the change in the fitness value. As the temperature decreases, the probability of accepting a worse solution gradually decreases. Therefore, the update strategy can be: in each iteration, the particle first updates its velocity and position according to the rules of PSO, and then judges whether to accept the new position according to the probability acceptance criterion of simulated annealing. If accepted, update the position of the particle; if not accepted, keep the original position unchanged.

[0097] Specifically, using the Simulated Annealing Particle Swarm Optimization (SAPSO) algorithm to optimize the Kernel Extreme Learning Machine (KELM) model, the main purpose is to find the optimal regularization parameter C and kernel parameter (such as the parameter σ of the radial basis kernel function) in the KELM model to improve the performance of the model. The following are the specific steps:

[0098] Step 1: Data preparation. Determine the historical unstable sequence corresponding to the historical power resource sequence and divide it into a training set and a test set.

[0099] Step 2: Initialize the SAPSO algorithm parameters, including: Particle swarm parameters: Set the size N of the particle swarm, and each particle represents a set of KELM model parameters (C and σ); initialize the position and velocity of each particle. Simulated annealing parameters: Set the initial temperature, temperature decay coefficient and termination temperature. Other parameters: Set the maximum number of iterations, individual learning factor and social learning factor.

[0100] Step 3: Define the fitness function. The fitness function is used to evaluate the performance of the KELM model represented by each particle. Commonly used fitness functions can be selected as the mean squared error (MSE), mean absolute error (MAE), etc. Here, MSE is taken as an example. For the parameters (C and σ) corresponding to each particle, construct a KELM model, train it using the training set data, and then make predictions on the test set. Calculate the MSE between the prediction results and the true values as the fitness value of this particle.

[0101] Step 4: Iteratively update the particle position and velocity. Calculate the individual extreme value and the global extreme value: At the initial moment, take the current position of each particle as its individual extreme value, and the position of the particle with the smallest fitness value as the global extreme value. Update the particle velocity and position: In each iteration, update the particle velocity and position according to the following formula:

[0102]

[0103] where w is the inertia weight, which provides a balance between local and global exploration (a larger w will promote global search, while a smaller w will promote local search), k is the number of iterations, d = 1, 2, 3…, m, positive numbers c1 and c2 are acceleration coefficients, r1 and r2 are uniformly distributed numbers in the range [0, 1], and N is the number of particles. The position of particle i in dimension d and iteration k is represented by denoted as.

[0104] After that, perform simulated annealing judgment. For the updated particle position, calculate its fitness value. If the fitness value of the new particle position is greater than the fitness value of the position before update, accept the new position; if the fitness value of the new particle position is less than the fitness value of the position before update, there is a probability of accepting the new position, where the probability is determined based on the current temperature. Then, update the individual extreme value and the global extreme value: If the new position is accepted and the fitness value of the new position is less than the fitness value corresponding to the individual extreme value, update the individual extreme value; if the fitness value of the individual extreme value is less than the fitness value of the global extreme value, determine the individual extreme value as the global extreme value.

[0105] Step 5: Judge the termination condition. Judge whether the termination condition is satisfied. Common termination conditions include reaching the maximum number of iterations or the temperature being lower than the termination temperature. If the termination condition is satisfied, end the iteration, output the global extreme value, and the corresponding parameters (C and σ); otherwise, return to Step 4 to continue the iteration.

[0106] Step 6: Construct the optimal KELM model. Construct a KELM model using the optimal parameters (C and σ) obtained by the SAPSO method. Use the training set data to train the optimal KELM model, and then make predictions on the test set to evaluate the performance of the model.

[0107] In this embodiment, the KELM model optimized by the SAPSO algorithm can improve the efficiency of training the KELM model and the prediction accuracy of the trained KELM model.

[0108] In an exemplary embodiment, the method further includes:

[0109] Based on a preset performance evaluation strategy, evaluate the target prediction result and the actual result corresponding to the target prediction result to obtain the prediction performance of the random time series model and the KELM model.

[0110] Among them, the performance evaluation strategy includes at least one of mean absolute error, mean absolute percentage error, and root mean square error.

[0111] Specifically, the server can use mean absolute error, mean absolute percentage error, and root mean square error to evaluate the prediction performance of the proposed model. The lower the error, the better the prediction performance of the model. The definition of the performance evaluation strategy is as follows:

[0112]

[0113]

[0114]

[0115] In the formula, Y i and y i respectively represent the actual value and the predicted value at time i, Y is the average actual value of the day, and n is the number of data samples.

[0116] In this embodiment, by using mean absolute error, mean absolute percentage error, and root mean square error to determine the prediction performance of the random time series model and the KELM model, it can provide data support for subsequent optimization of the random time series model and the KELM model.

[0117] As Figure 4 shown, the following combines a specific embodiment to detail the specific execution process of the above power resource prediction method. Specifically, it includes:

[0118] The first part, wavelet transform.

[0119] Wavelet transform is a mathematical tool for signal analysis. It can decompose a signal into different frequency components and analyze the performance of each component at different time points. Compared with the traditional Fourier transform, wavelet transform has better time-frequency localization ability and can analyze the signal in both the time domain and the frequency domain. Wavelet transform can be divided into two types: continuous wavelet transform (CWT) and discrete wavelet transform (DWT).

[0120] For the signal , the CWT is defined as:

[0121]

[0122] Where α is the scale parameter, τ is the translation parameter, ψ*(x) is the complex conjugate function, and ψ(x) is the mother wavelet.

[0123] The DWT is defined as:

[0124]

[0125] Where m is the scaling constant for the decomposition level and the integer n is the translation constant.

[0126] The fast wavelet transform algorithm (Mallat algorithm) uses low-pass and high-pass filters to extract irregular information from the original signal. The low-pass filter is called the scaling function and is used to analyze the low-frequency components; the high-pass filter is called the wavelet function and is used to analyze the high-frequency components. As Figure 5 shown, from the given signal F(t), an approximation sequence d and detail sequences (c1, c2, c3, c4) are obtained through the Mallat algorithm. The approximation sequence represents the low-frequency components and contains the trend information of the original signal; the detail sequences represent the high-frequency components and contain the characteristic information of the influencing factors related to the original signal. This model uses Daubechies5 as the mother wavelet to perform a five-level decomposition on the electricity price signal.

[0127] The second part, the ARMA time series model.

[0128] The ARMA model is a stochastic time series model based on the autocorrelation function and partial autocorrelation function, reflecting the internal connection of data between past and current activities. An ARMA(p, q) model of order p and q is defined as:

[0129]

[0130] Where is the ARMA sequence, the random term is an independent white noise sequence, is the autocorrelation coefficient, and θq is the moving average coefficient. If p = 0, the equation becomes a MA model of order q; if q = 0, it becomes an AR model of order p.

[0131] The ARMA sub-model is constructed according to the following process:

[0132] (1) Stationarity analysis: Analyze whether the data is a stationary sequence through the autocorrelation function and unit root test. If the autocorrelation coefficient decays exponentially and approaches zero, it indicates that the time series satisfies the stationarity assumption. (2) Identify the ARMA(p,q) model structure: Construct the ARMA function through the autocorrelation function ACF and partial autocorrelation function PACF. (3) Model parameter estimation and model construction. (4) Model testing, where the estimated residuals are white noise. (5) Predict future values.

[0133] Part three, the ARMA time series model.

[0134] The Extreme Learning Machine (ELM) is a single-hidden layer feedforward neural network (SLFN), which has a fast learning speed and good generalization ability. The advantage of ELM is that the parameters of the hidden layer do not need to be tuned, and the input weights and biases are randomly generated, so the calculation can be completed in a relatively short time. For N arbitrary samples , , , i = 1, 2…n. If the activation function of the hidden layer is g(x), then the output matrix T of the SLFN with L hidden nodes is:

[0135]

[0136] The above formula can be written as:

[0137]

[0138] In the formula, H is the output matrix of the hidden layer, the hidden layer threshold is b, and g is the activation function. It can be expressed as:

[0139]

[0140] The typical SLFN model can be expressed as:

[0141]

[0142] In the formula, β is the output weight between the hidden layer and the output layer, and G is the output function of the hidden node.

[0143] If the training set is , i = 1, 2,…, N, where , , 1, and are the same vectors, and the single-layer feedforward neural network with L hidden nodes can approximate these N samples with zero error, that is:

[0144]

[0145] Among them, O is the actual output value of the SLFN. If g, , b exist, then there is:

[0146]

[0147] Therefore, the matrix can also be expressed as:

[0148]

[0149] The input weights and hidden biases are randomly generated. The only unknown parameter is the output weight β, which can be solved by ordinary least squares.

[0150] The solution of the above formula is:

[0151]

[0152] In the formula, is the Moore-Penrose generalized inverse matrix of matrix H. Using this method, ELM can often obtain good generalization ability and fast learning speed. According to the orthogonal projection method and ridge regression theory, the output weight β can be calculated by adding a positive value 1 / λ as follows:

[0153]

[0154] The output function of ELM is:

[0155]

[0156] If the user does not know the feature mapping h(x), a kernel matrix can be defined for ELM, that is:

[0157]

[0158] The user does not need to know the hidden layer feature mapping h(x), and the function does not specify the number of hidden nodes L. The kernel k(u, v) that replaces h(x) and L is provided by the user. λ is the penalty factor. The stable kernel function replaces the random mapping of ELM, and the output weight is more stable. Therefore, the generalization ability of KELM is better than that of ELM. The kernel method was originally used in support vector machines and is widely applied to classification and regression problems, mapping the classification of nonlinear data in the original space to a high-dimensional space. This operation usually has a smaller computational burden than explicitly calculating coordinates. Different kernel functions have different learning abilities and generalization performances. Three types of kernel functions are used in this model: polynomial kernel, hyperbolic kernel, and Gaussian kernel.

[0159] Part 4, KELM optimized by the SAPSO algorithm.

[0160] The penalty factor and kernel parameter are two important parameters that affect the generalization ability of KELM. In the embodiments of this application, the SAPSO algorithm is adopted to adjust the penalty factor and kernel parameter of KELM during the process of training a neural network, so as to achieve stability and more effective regression performance. The particle swarm optimization algorithm PSO is an optimization method derived from the study of biological group behavior. In PSO, the solution to the optimization problem is represented as a particle in the search space. Each particle finds the optimal solution through the search space, and the particle position is updated by the velocity. Each particle has a fitness value determined by its objective function to evaluate the quality of the current position of the particle. The particle moves in the multi-dimensional search space at a random and adaptive speed to find a lower function value. The population X consists of N particles located in an m-dimensional search space. Each particle has two vectors, the position and the velocity , where i = 1, 2, 3, …, m. When substituting into the objective function, the corresponding fitness value will be calculated, which is the measurement standard for determining the particle mass. During the process of searching for the local optimal and global optimal positions, each particle will accelerate. Pi is the local optimal position, while Pj is the global optimal position. During the iteration process, each particle updates its velocity and position through Pi and Pj. According to the fitness, if the current position is better than the best position Pi, then Pi replaces the current position. Otherwise, Pi remains unchanged. The change of Pj is similar. This process will continue for more iterations until the maximum number of iterations is reached or the minimum error of the objective function is achieved. The velocity and position can be updated using the following formulas:

[0161]

[0162]

[0163] In the formula, w is the inertia weight, which provides a balance between local and global exploration (a larger w will promote global search, while a smaller w will promote local search), k is the number of iterations, d = 1, 2, 3, …, m, the positive numbers c1 and c2 are acceleration coefficients, r1 and r2 are uniformly distributed numbers in the range [0, 1], and N is the number of particles. The position of particle i in dimension d and iteration k is represented by. An adaptive inertia weight update strategy is adopted to balance the search ability of the algorithm:

[0164]

[0165] In the formula, w1 and w2 are the maximum value and the final value of the inertia weight respectively. K is the current number of iterations, and T is the maximum number of allowed iterations. This strategy can well balance global search and local search throughout the search process.

[0166] The SAPSO-KELM sub-model is constructed according to the following process:

[0167] (1) Initialize the size, velocity, and position of the particles; (2) Calculate the fitness value through the objective function; (3) Update the local best position Pi, the global best position Pj, the velocity and position of each particle according to the fitness value; (4) Repeat step (3) until the termination condition is met; (5) Output the optimal penalty factor and kernel parameter to KELM.

[0168] SAPSO-KELM has better generalization performance after using the optimal kernel parameter.

[0169] Part Five, prediction error measure.

[0170] The embodiments of this application use the mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE) to evaluate the prediction performance of the proposed model. A lower error metric indicates better prediction performance. The definitions are as follows:

[0171]

[0172]

[0173]

[0174] In the formula, and represent the actual value and predicted value at time i respectively, Y is the average actual value of the day, and n is the number of data samples.

[0175] In this embodiment, the model combines wavelet transform, kernel extreme learning machine (KELM) based on kernel method, and autoregressive moving average (ARMA), and uses self-adaptive particle swarm optimization (SAPSO) to search for the best kernel parameter of KELM. After decomposing the original sequence into a stationary sequence and a non-stationary sequence by wavelet, the ARMA model is used to predict the stationary sequence as a new input set, while the non-stationary sequence is predicted by the SAPSO-KELM model. This hybrid model combines the advantages of each sub-model, has the prediction ability for both linear and non-linear electricity price sequences, and can effectively improve the accuracy of day-ahead electricity price prediction.

[0176] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, at least some of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least some of the steps or stages in other steps or other steps.

[0177] Based on the same inventive concept, an embodiment of the present application further provides a power resource prediction device for implementing the power resource prediction method described above. The solution provided by this device to solve the problem is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the power resource prediction device provided below can refer to the limitations on the power resource prediction method in the above text, and will not be repeated here.

[0178] In an exemplary embodiment, as Figure 6 shown, a power resource prediction device 600 is provided, including: a sequence decomposition module 601, a first prediction module 602, a second prediction module 603, and a result determination module 604, where:

[0179] The sequence decomposition module 601 is configured to decompose the power resource sequence based on a wavelet transform strategy to obtain a stationary sequence and a non-stationary sequence;

[0180] The first prediction module 602 is configured to predict the stationary sequence based on a preset random time series model to obtain a first prediction result; the random time series model includes an AR sub-model and an MA sub-model; the AR sub-model and the MA sub-model are determined based on the autocorrelation function and the partial autocorrelation function corresponding to the historical power resource sequence;

[0181] The second prediction module 603 is configured to predict the non-stationary sequence based on a preset kernel extreme learning machine KELM model to obtain a second prediction result; the KELM model is trained based on the historical power resource sequence and a simulated annealing particle swarm optimization SAPSO strategy;

[0182] The result determination module 604 is configured to obtain a target prediction result corresponding to the power resource sequence based on the first prediction result and the second prediction result.

[0183] Further, the wavelet transform strategy includes a fast algorithm for multi-resolution analysis and a sequence decomposition module 601, which is specifically configured to: filter the power resource sequence based on the low-pass filter in the fast algorithm for multi-resolution analysis to obtain an approximation sequence; the approximation sequence represents the trend information of the power resource sequence; filter the power resource sequence based on the high-pass filter in the fast algorithm for multi-resolution analysis to obtain a detail sequence; the detail sequence represents the characteristic of influencing factors related to the power resource sequence; based on the autocorrelation function and the unit root test strategy of the power resource sequence, perform stationarity analysis on the approximation sequence and the detail sequence respectively to determine the sequence types of the approximation sequence and the detail sequence; the sequence types include stationary sequences and non-stationary sequences.

[0184] Further, the first prediction module 602 is specifically configured to: input each element into the AR sub-model and the MA sub-model in the random time series model according to the time of each element in the stationary sequence to obtain the prediction result of the AR sub-model and the prediction result of the MA sub-model; add the prediction result of the AR sub-model and the prediction result of the MA sub-model to obtain the first prediction result corresponding to the stationary sequence.

[0185] Further, the second prediction module 603 is specifically configured to: input the non-stationary sequence into the kernel mapping layer of the KELM model to obtain the kernel vector corresponding to the non-stationary sequence; perform weighted summation on the output weight and the kernel vector of the KELM model to obtain the second prediction result corresponding to the non-stationary sequence.

[0186] Further, the device further includes a first training module, which is specifically configured to: obtain the historical stationary sequence in the historical power resource sequence; determine the ARMA model based on the autocorrelation function and the partial autocorrelation function corresponding to the historical stationary sequence; estimate the parameters in the ARMA model to obtain an updated ARMA model, and determine the white noise test result of the updated ARMA model; if the residual sequence in the white noise test result is not a white noise sequence, then return to execute the step of estimating the parameters in the ARMA model to obtain an updated ARMA model until the residual sequence in the white noise test result is a white noise sequence, then determine the updated ARMA model as the random time series model.

[0187] Further, the device further includes a second training module, which is specifically configured to: determine the particles in the SAPSO strategy as the regularization parameter and the kernel parameter of the KELM model; determine the fitness values corresponding to the regularization parameter and the kernel parameter based on the fitness function; update the local best position, the global best position, the velocity and the position of each particle according to the fitness values until the training termination condition is met to obtain the optimal regularization parameter and kernel parameter; determine the trained KELM model based on the optimal regularization parameter and kernel parameter.

[0188] Further, the device further includes a performance evaluation module, specifically configured to: evaluate the target prediction result and the actual result corresponding to the target prediction result based on a preset performance evaluation strategy, so as to obtain the prediction performance of the random time series model and the KELM model; the performance evaluation strategy includes at least one of mean absolute error, mean absolute percentage error, and root mean square error.

[0189] Each module in the above power resource prediction device can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in the processor in the computer device in hardware form or be independent of the processor, or can be stored in the memory in the computer device in software form, so that the processor can call and execute the operations corresponding to the above respective modules.

[0190] In an exemplary embodiment, a computer device is provided. The computer device can be a server, and its internal structure diagram can be as Figure 7 shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store historical power resource sequences. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through a network connection. When the computer program is executed by the processor, it implements a power resource prediction method.

[0191] Those skilled in the art can understand that Figure 7 the structure shown in

[0192] is only a block diagram of some structures related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0193] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps in the foregoing method embodiments are implemented.

[0194] In one embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the foregoing method embodiments are implemented.

[0195] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.

[0196] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memories can include read-only memory (ROM), magnetic tapes, floppy disks, flash memories, optical memories, high-density embedded non-volatile memories, resistive random access memories (ReRAM), magnetoresistive random access memories (MRAM), ferroelectric random access memories (FRAM), phase change memories (PCM), graphene memories, etc. Volatile memories can include random access memory (RAM) or external cache memories, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processors, graphics processors, digital signal processors, programmable logics, data processing logics based on quantum computing, etc., without limitation.

[0197] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0198] The above-described embodiments merely represent several implementation manners of the present application. Their descriptions are relatively specific and detailed, but they should not be construed as limiting the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A power resource prediction method, characterized in that, The method includes: Decompose the power resource sequence based on the wavelet transform strategy to obtain a stationary sequence and a non-stationary sequence; Predict the stationary sequence based on a preset random time series model to obtain a first prediction result; the random time series model includes an AR sub-model and an MA sub-model; the AR sub-model and the MA sub-model are determined based on the autocorrelation function and the partial autocorrelation function corresponding to the historical power resource sequence; Predict the non-stationary sequence based on a preset Kernel Extreme Learning Machine (KELM) model to obtain a second prediction result; the KELM model is trained based on the historical power resource sequence and the Simulated Annealing Particle Swarm Optimization (SAPSO) strategy; Obtain the target prediction result corresponding to the power resource sequence based on the first prediction result and the second prediction result.

2. The method according to claim 1, wherein The wavelet transform strategy includes a multi-resolution analysis fast algorithm; the decomposing the power resource sequence based on the wavelet transform strategy to obtain a stationary sequence and a non-stationary sequence includes: Filter the power resource sequence based on the low-pass filter in the multi-resolution analysis fast algorithm to obtain an approximation sequence; the approximation sequence represents the trend information of the power resource sequence; Filter the power resource sequence based on the high-pass filter in the multi-resolution analysis fast algorithm to obtain a detail sequence; the detail sequence represents the influencing factor characteristics related to the power resource sequence; Conduct stationary analysis on the approximation sequence and the detail sequence respectively based on the autocorrelation function and the unit root test strategy of the power resource sequence to determine the sequence types of the approximation sequence and the detail sequence; the sequence types include a stationary sequence and a non-stationary sequence.

3. The method according to claim 1, characterized in that, The predicting the stationary sequence based on a preset random time series model to obtain a first prediction result includes: Input each element of the stationary sequence into the AR sub-model and the MA sub-model in the random time series model according to the time of each element in the stationary sequence to obtain the prediction result of the AR sub-model and the prediction result of the MA sub-model; Add the prediction result of the AR sub-model and the prediction result of the MA sub-model to obtain the first prediction result corresponding to the stationary sequence.

4. The method according to claim 1, wherein The predicting the non-stationary sequence based on a preset Kernel Extreme Learning Machine (KELM) model to obtain a second prediction result includes: Input the non-stationary sequence into the kernel mapping layer of the KELM model to obtain the kernel vector corresponding to the non-stationary sequence; Perform weighted summation on the output weight of the KELM model and the kernel vector to obtain the second prediction result corresponding to the non-stationary sequence.

5. The method according to claim 1, wherein The method further includes: Obtain the historical stationary sequence in the historical power resource sequence; Determine an ARMA model based on the autocorrelation function and the partial autocorrelation function corresponding to the historical stationary sequence; Estimate each parameter in the ARMA model to obtain an updated ARMA model, and determine the white noise test result of the updated ARMA model; If the residual sequence in the white noise test result is not a white noise sequence, return to execute the step of determining each parameter in the estimated ARMA model to obtain an updated ARMA model until the residual sequence in the white noise test result is a white noise sequence, and then determine the updated ARMA model as the random time series model.

6. The method according to claim 1, wherein The method further includes: Determining the particles in the SAPSO strategy as the regularization parameter and kernel parameter of the KELM model; Determining the fitness value corresponding to the regularization parameter and kernel parameter based on the fitness function; Updating the local best position, global best position, the velocity and position of each particle according to the fitness value until the training termination condition is satisfied to obtain the optimal regularization parameter and kernel parameter; Determining a trained KELM model based on the optimal regularization parameter and kernel parameter.

7. The method according to claim 1, wherein The method further includes: Evaluating the target prediction result and the actual result corresponding to the target prediction result based on a preset performance evaluation strategy to obtain the prediction performance of the random time series model and the KELM model; the performance evaluation strategy includes at least one of mean absolute error, mean absolute percentage error and root mean square error.

8. A power resource prediction device, characterized in that, The device includes: A sequence decomposition module, configured to decompose the power resource sequence based on a wavelet transform strategy to obtain a stationary sequence and a non-stationary sequence; A first prediction module, configured to predict the stationary sequence based on a preset random time series model to obtain a first prediction result; the random time series model includes an AR sub-model and an MA sub-model; the AR sub-model and the MA sub-model are determined based on the autocorrelation function and partial autocorrelation function corresponding to the historical power resource sequence; A second prediction module, configured to predict the non-stationary sequence based on a preset kernel extreme learning machine (KELM) model to obtain a second prediction result; the KELM model is trained based on the historical power resource sequence and a simulated annealing particle swarm optimization (SAPSO) strategy; A result determination module, configured to obtain a target prediction result corresponding to the power resource sequence based on the first prediction result and the second prediction result.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images