A load prediction method and system based on big data

Abstract

The application discloses a load prediction method and system based on big data, and relates to the technical field of power grid load prediction.The method comprises the following steps: based on a least squares support vector machine, pre-processing load historical data; through decomposing the load historical data with reduced complexity, obtaining a zero-crossing rate and sample entropy, and determining the multi-frequency components of the load data; through a preset hybrid algorithm and in combination with the multi-frequency components of the load data, training a multi-factor weighted combination analysis model; based on the multi-factor weighted combination analysis model and according to an improved grey wolf algorithm, determining the weight of the load prediction result of each prediction factor module; and weighting and combining the load prediction results of each prediction factor module to obtain a final load prediction result.The application improves the robustness of the model, dynamically updates the model output result, optimizes the weight proportion among the factor modules, and improves the adaptability of load prediction in a complex environment.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of power grid load prediction, and particularly relates to a load prediction method and system based on big data. BACKGROUND

[0002] With the rapid development of new power systems, the power consumption load scenarios within the power system are becoming increasingly diversified, which brings new challenges and opportunities to load prediction and management. New power systems include industrial loads, commercial loads, residential loads, agricultural loads, and emerging electric vehicle loads, data center loads, etc. Each type of load has its unique power consumption characteristics and variation rules. With the access of renewable energy and the development of distributed energy, the volatility of load becomes more complex, which puts higher requirements on the accuracy of load prediction. At the same time, the application of smart grid and demand response technology makes the load no longer a passive acceptor, but can interact with the power system. Therefore, the enhanced load interaction requires load prediction to consider changes in user behavior and market demand. In order to cope with these challenges, it is necessary to apply corresponding intelligent prediction methods for different types of load prediction scenarios to obtain more accurate and efficient load prediction results, thereby significantly improving the load management capability of new power systems.

[0003] The patent with the authorized publication number CN112100867B discloses a power grid load prediction method, which includes: correcting the parameters of daily maximum load data and daily average temperature data according to first and second correction coefficients, respectively, to obtain a corrected GM(1,2) model; performing cumulative subtraction operation on the corrected GM(1,2) model to obtain a power grid daily load prediction model, and using the power grid daily load prediction model to predict the power grid of the next day. The above-mentioned power grid load prediction method discloses the prediction of the power grid daily load prediction model through historical data. However, in the input process of a large amount of data, the prediction accuracy of the traditional model for different types of data is poor, resulting in high demand for computing resources.

[0004] At present, there is no effective solution to the problems in the related art. SUMMARY

[0005] In order to solve the problem in the related art that the prediction accuracy of the traditional model for different types of data is poor in the input process of a large amount of data, resulting in high demand for computing resources, the present application proposes a load prediction method based on big data to overcome the above technical problems existing in the prior art.

[0006] To this end, the specific technical solutions adopted by the present application are as follows:

[0007] According to one aspect of the present application, a load prediction method based on big data is provided, which includes:

[0008] S1, based on least squares support vector machine, the load history data is nonlinearly transformed, and the Lagrange function form is defined, and the vector regression function expression is solved to reduce the complexity of the load history data.

[0009] S2, by decomposing the load history data with reduced complexity, the zero-crossing rate and sample entropy are obtained; the multi-frequency component of the load data is determined by using the zero-crossing rate and sample entropy.

[0010] S3, through the preset hybrid algorithm, and combining the multi-frequency component of the load data, a multi-module weighted combination analysis model is trained.

[0011] S4, based on the prediction performance, model complexity and confidence of each prediction module in the multi-module weighted combination analysis model, and according to the improved grey wolf algorithm, the weight of the load prediction result of each prediction module is determined; the load prediction results of each prediction module are combined to obtain the final load prediction result.

[0012] Further, based on the least squares support vector machine, the load history data is nonlinearly transformed, and the Lagrange function form is defined, and the vector regression function expression is solved to reduce the complexity of the load history data, including:

[0013] Collect the training sample set corresponding to the load history data, and use nonlinear variables to convert the input vectors in the training sample set into nonlinear problems to obtain the function of the least squares support vector machine; according to the function of the least squares support vector machine, define the Lagrange function, and input the space function after conditional optimization and solving to form the vector regression function expression:

[0014]

[0015] In the formula, f3(x) represents the vector regression function, i represents any constant not equal to zero, x i represents the i-th input, b represents the bias value of the adjacent uncertain regression function, K represents the space function, a i represents the sample actual value, and x represents the new input variable to be predicted.

[0016] Further, by decomposing the load history data with reduced complexity, the zero-crossing rate and sample entropy are obtained, including:

[0017] The number of iterations of the complete ensemble empirical mode decomposition with adaptive noise and the added noise amplitude are configured; the history data with reduced complexity is decomposed into intrinsic mode function components and residual components by the complete ensemble empirical mode decomposition with adaptive noise; the zero-crossing rate and sample entropy are obtained from the intrinsic mode function components.

[0018] Further, the multiple frequency components of the load data are determined by using the zero-crossing rate and the sample entropy, and the multiple frequency components include:

[0019] The high frequency component, the medium frequency component and the low frequency component of the load data are determined by using the zero-crossing rate and the sample entropy, wherein the high frequency component contains the influence of noise and sudden events, the medium frequency component contains daily or weekly periodicity, and the low frequency component contains long-term trend and seasonality.

[0020] Further, the multi-cause weighted combination analysis model is trained by using a preset hybrid algorithm and in combination with the multiple frequency components of the load data, and the multi-cause weighted combination analysis model includes:

[0021] The wavelet transform and convolution neural network model are trained by using the wavelet transform and the convolution neural network and in combination with the high frequency component, the recurrent neural network model is trained by using the empirical mode decomposition and the recurrent neural network and in combination with the medium frequency component, and the random forest model is trained by using the Fourier transform and the random forest and in combination with the low frequency component, the wavelet transform and convolution neural network model, the recurrent neural network model and the random forest model are combined to obtain the multi-cause weighted combination analysis model.

[0022] Further, the weights of the load prediction results of the prediction cause modules in the multi-cause weighted combination analysis model are determined based on the prediction performance, the model complexity and the confidence of each prediction cause module and in accordance with the improved grey wolf algorithm, and the weights of the load prediction results of the prediction cause modules in the multi-cause weighted combination analysis model include:

[0023] The initial weights of the prediction cause modules in the multi-cause weighted combination analysis model are allocated, the prediction performance index, the model complexity and the confidence of each prediction cause module are obtained, the iteration update of the improved grey wolf algorithm is used in combination with the fitness function including the prediction performance index, the model complexity and the confidence to complete the weight optimization of the output results of the prediction cause modules, and the prediction cause modules include the wavelet transform and convolution neural network model, the recurrent neural network model and the random forest model.

[0024] Further, the iteration update of the improved grey wolf algorithm is used in combination with the fitness function including the prediction performance index, the model complexity and the confidence to complete the weight optimization of the output results of the prediction cause modules, and the iteration update of the improved grey wolf algorithm includes:

[0025] The parameters of the improved grey wolf algorithm are configured, the fitness value of each group of weight vectors is calculated according to the prediction performance index, the model complexity and the confidence of each prediction cause module and the corresponding weight vector, the position of the wolf group is iteratively updated according to the preset update rule to find the optimal weight vector, and the weight vector with the highest fitness value is selected as the optimal weight.

[0026] Further, the fitness function is:

[0027] Fitness = alpha * prediction performance + beta * model complexity penalty + gamma * confidence adjustment fitness; in the formula, alpha, beta and gamma respectively represent the weight vectors corresponding to prediction performance, model complexity penalty and confidence adjustment fitness.

[0028] Further, the prediction performance indicators of each predictive meme module are obtained, including:

[0029] The prediction performance indicators of each predictive meme module on the test set are calculated by mean square error.

[0030] According to another aspect of the present application, a big data-based load prediction system is also provided, comprising:

[0031] The data preprocessing module is configured to perform nonlinear transformation on the load historical data based on the least squares support vector machine, define a Lagrange function form, and solve to obtain a vector regression function expression, so as to reduce the complexity of the load historical data.

[0032] The data input matching module is configured to obtain the zero-crossing rate and sample entropy by decomposing the load historical data with reduced complexity, and determine the multi-frequency components of the load data by using the zero-crossing rate and sample entropy.

[0033] The model training module is configured to train a multi-meme weighted combination analysis model by using a preset hybrid algorithm and in combination with the multi-frequency components of the load data.

[0034] The weight combination module is configured to determine the weights of the load prediction results of each predictive meme module based on the prediction performance, model complexity and confidence of each predictive meme module in the multi-meme weighted combination analysis model and according to the improved grey wolf algorithm, and combine the load prediction results of each predictive meme module to obtain the final load prediction result.

[0035] As described above, due to the adoption of the above technical solutions, the present application has the following beneficial effects:

[0036] In the present application, by using the combination prediction method, different types of power grid load prediction data can be matched with the characteristics of various meme modules, the model robustness can be improved, the least squares support vector machine is used for fast anomaly repair, the linear equation solving characteristics are used to effectively reduce the calculation complexity, the heterogeneous model cluster is constructed, the algorithm characteristics and data frequency domain characteristics are accurately matched, the challenges of data heterogeneity and prediction demand diversification in the big data environment can be adapted, and the model output result can be dynamically updated to optimize the weight proportion between the meme modules, so that the adaptability of load prediction in complex environment can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0038] Figure 1 is a flowchart of a load prediction method based on big data according to an embodiment of the present application;

[0039] Figure 2 is a flowchart in actual process according to the present application. DETAILED DESCRIPTION

[0040] In order to further illustrate the embodiments, the present application provides drawings which are part of the disclosure of the present application, mainly used to illustrate the embodiments, and can be used to explain the operating principle of the embodiments in conjunction with the related description of the specification. With reference to these contents, those skilled in the art should understand other possible embodiments and advantages of the present application. The components in the drawings are not drawn to scale, and similar component symbols are usually used to represent similar components.

[0041] According to an embodiment of the present application, a load prediction method and system based on big data are provided.

[0042] The present application will be further described in conjunction with the drawings and specific embodiments. As shown in the drawings, according to an embodiment of the present application, a load prediction method based on big data is provided, comprising: Figure 1

[0043] S1, based on least squares support vector machine, nonlinearly transforming load historical data, defining Lagrange function form, and solving to obtain vector regression function expression, so as to reduce the complexity of load historical data.

[0044] S2, by decomposing the load historical data with reduced complexity, obtaining zero-crossing rate and sample entropy; using the zero-crossing rate and sample entropy, determining the multi-frequency components of the load data.

[0045] S3, by using a preset hybrid algorithm and combining the multi-frequency components of the load data, a multi-cause weighted combination analysis model is trained.

[0046] S4, based on the prediction performance, model complexity and confidence of each prediction cause module in the multi-cause weighted combination analysis model, and according to the improved grey wolf algorithm, the weight of the load prediction result of each prediction cause module is determined; the load prediction results of each prediction cause module are weighted and combined to obtain the final load prediction result.

[0047] ​In one embodiment, based on the least squares support vector machine, the load history data is nonlinearly transformed, and a Lagrange function form is defined, and a vector regression function expression is obtained by solving, so as to reduce the complexity of the load history data, including:

[0048] The training sample set corresponding to the load history data is collected, and the input vector in the training sample set is transformed into a nonlinear problem by using a nonlinear variable, so as to obtain a function of the least squares support vector machine; according to the function of the least squares support vector machine, a Lagrange function is defined, and after the conditional optimization and the input space function are solved, a vector regression function expression is formed:

[0049]

[0050] In the formula, f3(x) represents the vector regression function, i represents an arbitrary constant other than zero, x i represents the i th input quantity, b represents a bias value of the adjacent uncertain regression function, K represents a space function, a i represents a sample actual value, and x represents a new input variable to be predicted, that is, an arbitrary point at which the regression function value needs to be evaluated.

[0051] In one embodiment, the zero-crossing rate and the sample entropy are obtained by decomposing the load history data with reduced complexity, including:

[0052] The number of iterations of the complete ensemble empirical mode decomposition with adaptive noise and the amplitude of the added noise are configured; the history data with reduced complexity is decomposed into intrinsic mode function components and residual components by the complete ensemble empirical mode decomposition with adaptive noise; the zero-crossing rate and the sample entropy are obtained from the intrinsic mode function components.

[0053] In one embodiment, the multi-frequency component of the load data is determined by using the zero-crossing rate and the sample entropy, including:

[0054] The high-frequency component, the medium-frequency component and the low-frequency component of the load data are determined by using the zero-crossing rate and the sample entropy; wherein the high-frequency component contains the influence of noise and sudden events, the medium-frequency component contains daily periodicity or weekly periodicity, and the low-frequency component contains long-term trend and seasonality.

[0055] In one embodiment, a multi-modal weighted combination analysis model is trained by using a preset hybrid algorithm and in combination with the multi-frequency component of the load data, including:

[0056] The wavelet transform and the convolution neural network are combined with the high-frequency component to train a wavelet transform and convolution neural network model to perform high-frequency component prediction; the empirical mode decomposition and the recurrent neural network are combined with the medium-frequency component to train a recurrent neural network model to perform medium-frequency component prediction; the Fourier transform and the random forest are combined with the low-frequency component to train a random forest model to perform low-frequency component prediction; and the wavelet transform and convolution neural network model, the recurrent neural network model and the random forest model are combined to obtain a multi-module weighted combination analysis model.

[0057] In one embodiment, based on the prediction performance, model complexity and confidence of each prediction module in the multi-module weighted combination analysis model, and according to the improved grey wolf algorithm, the weight of the load prediction result of each prediction module is determined, including:

[0058] allocating initial weights of each prediction module in the multi-module weighted combination analysis model; obtaining the prediction performance index, model complexity and confidence of each prediction module; using the iterative update of the improved grey wolf algorithm and combining the fitness function including the prediction performance index, model complexity and confidence to complete the weight optimization of the output result of each prediction module; wherein the prediction module includes the wavelet transform and convolution neural network model, the recurrent neural network model and the random forest model.

[0059] In one embodiment, using the iterative update of the improved grey wolf algorithm and combining the fitness function including the prediction performance index, model complexity and confidence to complete the weight optimization of the output result of each prediction module includes:

[0060] configuring parameters of the improved grey wolf algorithm, calculating the fitness value of each group of weight vectors according to the prediction performance index, model complexity and confidence of each prediction module, and the corresponding weight vector; iteratively updating the position of the wolf group according to the preset update rule to find the optimal weight vector; and selecting the weight vector with the highest fitness value as the optimal weight.

[0061] In one embodiment, the fitness function is:

[0062] Fitness = a·prediction performance + β·model complexity penalty + γ·confidence adjustment fitness; wherein a, β and γ represent the weight vectors corresponding to the prediction performance, model complexity penalty and confidence adjustment fitness, respectively.

[0063] In one embodiment, obtaining the prediction performance index of each prediction module includes:

[0064] The prediction performance index of each prediction module on the test set is calculated by the mean square error.

[0065] According to another embodiment of the present application, a big data-based load forecasting system is also provided, comprising:

[0066] A data preprocessing module is configured to perform nonlinear transformation on the load historical data based on a least squares support vector machine, define a Lagrange function form, and solve a vector regression function expression to reduce the complexity of the load historical data.

[0067] A data input matching module is configured to obtain a zero-crossing rate and sample entropy by decomposing the load historical data with reduced complexity, and determine the multi-frequency components of the load data using the zero-crossing rate and sample entropy.

[0068] A model training module is configured to train a multi-factor weighted combination analysis model using a preset hybrid algorithm and in combination with the multi-frequency components of the load data.

[0069] A weight combination module is configured to determine the weights of the load forecasting results of each prediction factor module based on the prediction performance, model complexity, and confidence of each prediction factor module in the multi-factor weighted combination analysis model, and according to an improved grey wolf algorithm, and combine the load forecasting results of each prediction factor module to obtain a final load forecasting result.

[0070] In order to facilitate the understanding of the above technical solutions of the present application, the working principle of the present application in the actual process will be described in detail below.

[0071] As shown in Figure 2 , historical data is obtained, and data preprocessing is performed on the historical data. A multi-factor weighted combination analysis model is established, a plurality of prediction models are trained using a preset algorithm, and each prediction factor module is established based on the plurality of prediction models. The preprocessed historical data is sequentially decomposed to obtain historical data input sequences matching the data processing characteristics of each prediction factor module, and the input sequences are input into each prediction factor module. Each prediction factor module outputs a load forecasting result according to the characteristics of the model. The weights of the load forecasting results of each prediction factor module are assigned, and the load forecasting results of each prediction factor module are combined to obtain a final load forecasting result.

[0072] Data preprocessing on historical data includes:

[0073] Optimization of time series data specifically includes preprocessing power load data using a least squares support vector machine:

[0074] Collecting and setting a training sample set:

[0075] C={x i ,y i}|i=1,2,…,l|;

[0076] wherein C represents a preset result, xi represents input quantity, y i represents output quantity, l represents data length, i represents any constant not equal to zero. A certain training set function is designated as f(x) = wx i +b; w represents a vector perpendicular to a hyperplane, and b represents a bias value adjacent to an uncertain regression function.

[0077] Using nonlinear variable The input vector is converted into a nonlinear problem, and the function of the least squares support vector machine is:

[0078]

[0079] wherein γ represents a positive parameter, and e represents a sample deviation. According to the function definition of the least squares support vector machine according to the above formula, the Lagrange function L(w, b, e, a) is defined, and the expression is:

[0080]

[0081]

[0082] wherein, represents a sample actual value, a and b both represent real numbers, represents a nonlinear transformation.

[0083] The conditional optimization obtains the expression:

[0084]

[0085] Finally, the following is obtained:

[0086]

[0087] wherein E represents a weighted mean square error, and Z represents a set of nonlinear transformations,

[0088] The vector regression function is solved, and the expression is:

[0089]

[0090] Through the input space function K(x, x i ), the regression function expression is formed as:

[0091]

[0092] Through the above processing of historical abnormal data, the change index of non-missing data is input into the model (multi-factor weighted combination analysis model). The purpose is to reduce the data complexity of data preprocessing.

[0093] The preprocessed historical data is subjected to sequence decomposition, which comprises:

[0094] Initialize parameters, set the number of iterations of the complete ensemble empirical mode decomposition with adaptive noise, and the amplitude of the added noise.

[0095] The original load sequence is decomposed into several intrinsic mode function components and a residual component by the complete ensemble empirical mode decomposition with adaptive noise, each intrinsic mode function component representing a different frequency component in the original sequence. The residual component is an explanation of the function component.

[0096] The intrinsic mode function components are classified by zero-crossing rate and sample entropy. The zero-crossing rate represents the number of times the signal passes through zero in a unit of time, reflecting the frequency characteristics of the signal; sample entropy is used to measure the complexity of time series, reflecting the randomness and regularity of the signal.

[0097] According to the characteristics of the load data signal (zero-crossing rate and sample entropy), high-frequency components, medium-frequency components and low-frequency components are determined; the high-frequency components, medium-frequency components and low-frequency components are analyzed by characteristic classification of each prediction module. The high-frequency components contain the influence of noise and sudden events; the medium-frequency components contain daily or weekly periodicity; the low-frequency components contain long-term trends and seasonality.

[0098] The multi-module weighted combination analysis model trains multiple prediction models through a preset algorithm, which specifically comprises:

[0099] Through wavelet transform-convolution neural network, the high-frequency components are subjected to wavelet transform to extract time-frequency features, a convolution neural network model is constructed, the features after wavelet transform are input, and the wavelet transform-convolution neural network model is trained for high-frequency component prediction.

[0100] Through empirical mode decomposition-cyclic neural network, the medium-frequency components are subjected to empirical mode decomposition to extract main modes, a cyclic neural network model is constructed, the modes after empirical mode decomposition are input, and the cyclic neural network model is trained for medium-frequency component prediction.

[0101] Through Fourier transform-random forest, the low-frequency components are subjected to Fourier transform to extract frequency domain features, a random forest model is constructed, the features after Fourier transform are input, and the random forest model is trained for low-frequency component prediction.

[0102] The weights of the load prediction results of each prediction module are assigned, and the weights are determined according to the improved grey wolf algorithm based on the prediction performance of the model on historical data, the complexity of the model and the confidence factors, which specifically comprises:

[0103] A weight vector is defined for each model to represent its contribution to the final prediction, and the weight vector satisfies the constraint condition that the sum of all weights is 1.

[0104] Set parameters: including population size, maximum number of iterations, initial value of control parameters and decay rate.

[0105] Divide historical data into training set and test set, model training and prediction, respectively through each prediction module to train the training set, and predict on the test set.

[0106] Performance index calculation, calculate the prediction performance index of each model on the test set by mean square error.

[0107] Prediction performance, the prediction performance index of the model on the test set is used as a component of the fitness function.

[0108] Based on the model complexity as the penalty term of fitness function, according to the confidence [confidence is calculated by selecting appropriate method according to task demand (such as classification, regression) and computing resource] or uncertainty estimation of model, adjust its weight in fitness function, formula is:

[0109] Fitness = a prediction performance + beta model complexity penalty + gamma confidence adjustment fitness.

[0110] Where, a, beta, gamma are preset weight coefficients, the weight is optimized by grey wolf algorithm. Model complexity penalty is negative.

[0111] Population initialization: randomly generate a set of initial weight vectors as the initial position of grey wolf population.

[0112] Fitness evaluation: calculate the fitness value of each weight vector (a, beta, gamma).

[0113] Wolf population update: according to the update rule of control parameters, iteratively update the position of wolf population to find the optimal weight vector.

[0114] Alpha wolf is the current optimal solution; Beta wolf and delta wolf are suboptimal solution; other wolves are updated according to the position of alpha, beta and delta wolves.

[0115] Parameter update with iteration, update the value of control parameters, and after optimization, select the weight vector with the highest fitness value as the optimal weight, complete the weight optimization of the output results of each prediction module, and get the load prediction result through the optimized weight.

[0116] A load prediction system based on big data, comprising:

[0117] A historical data collection module for collecting historical data of power grid load through a data collection terminal.

[0118] The data preprocessing module is used for preprocessing the power load data to filter out abnormal data, reduce errors and improve prediction accuracy.

[0119] The multi-module weighted combination analysis model has multiple prediction module modules, the multiple prediction module modules are trained by different algorithms, and the multiple prediction module modules are used for analyzing the preprocessed historical data.

[0120] The data input matching module is used for processing historical data into data sequences matching different prediction module modules.

[0121] The weight combination module is used for weighting and combining the load prediction results output by the prediction module modules to obtain the final load prediction effect.

[0122] The multiple prediction module modules are respectively:

[0123] The wavelet transform-convolution neural network prediction module is used for wavelet transform-convolution neural network, wavelet transform of high-frequency components, extraction of time-frequency characteristics, construction of convolution neural network model, input of wavelet transformed features, and training of wavelet transform-convolution neural network model for high-frequency component prediction.

[0124] The empirical mode decomposition-cyclic neural network prediction module is used for empirical mode decomposition-cyclic neural network, empirical mode decomposition of medium-frequency components, extraction of main modes, construction of cyclic neural network model, input of empirical mode decomposition modes, and training of cyclic neural network model for medium-frequency component prediction.

[0125] The Fourier transform-random forest prediction module is used for Fourier transform-random forest of low-frequency components, extraction of frequency domain features, construction of random forest model, input of Fourier transformed features, and training of random forest model for low-frequency component prediction.

[0126] In summary, by combining the prediction method, the matching characteristics of different types of power grid load prediction data can be realized by using multiple module modules, the model robustness can be improved, the least squares support vector machine-based fast anomaly repair can effectively reduce the computational complexity by using its linear equation solving characteristics, the heterogeneous model cluster can be constructed, the algorithm characteristics and data frequency domain features can be accurately matched, the challenges of data heterogeneity and prediction demand diversification in the big data environment can be adapted, and the weight proportion between the module modules can be dynamically updated to optimize the model output results, and the adaptability of load prediction in complex environment can be improved.

[0127] The above only describes the preferred embodiments of the present application and does not limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A big data based load forecasting method, characterized in that, The method comprises the following steps: S1, based on the least squares support vector machine, the load history data is nonlinearly transformed, and the Lagrange function form is defined, and the vector regression function expression is obtained by solving, so as to reduce the complexity of the load history data; S2, the zero-crossing rate and sample entropy are obtained by decomposing the load history data with reduced complexity; the multi-frequency component of the load data is determined by using the zero-crossing rate and sample entropy; S3, through the preset hybrid algorithm, and combining the multi-frequency component of the load data, a multi-module weighted combination analysis model is trained, which specifically comprises: Through wavelet transform and convolution neural network, and combining high-frequency components, the wavelet transform and convolution neural network model is trained to predict high-frequency components; Through empirical mode decomposition and recurrent neural network, and combining the medium-frequency component, the recurrent neural network model is trained to predict the medium-frequency component; through Fourier transform and random forest, and combining the low-frequency component, the random forest model is trained to predict the low-frequency component; the wavelet transform and convolution neural network model, the recurrent neural network model and the random forest model are combined to obtain the multi-module weighted combination analysis model; S4, based on the prediction performance, model complexity and confidence of each prediction module in the multi-module weighted combination analysis model, and according to the improved grey wolf algorithm, the weight of the load prediction result of each prediction module is determined; the load prediction results of each prediction module are weighted and combined to obtain the final load prediction result.

2. The load forecasting method based on big data according to claim 1, characterized in that, The least squares support vector machine is used to transform the load history data into a nonlinear function, and the Lagrange function form is defined, and the vector regression function expression is obtained by solving, so as to reduce the complexity of the load history data, which comprises: Collecting the training sample set corresponding to the load history data, and using nonlinear variables to transform the input vectors in the training sample set into nonlinear problems to obtain the function of the least squares support vector machine; According to the function of the least squares support vector machine, the Lagrange function is defined, and the vector regression function expression is formed by condition optimization and solving the input space function: ； wherein, Vector regression function, represents an arbitrary constant not equal to zero, x i represents the i-th input quantity, b represents a bias value adjacent to the uncertain regression function, represents a spatial function, represents a sample actual value, represents a new input variable to be predicted regression.

3. The load prediction method based on big data according to claim 1, characterized in that, The complexity-reduced load history data is decomposed to obtain the zero-crossing rate and sample entropy, which comprises: Configuring the iteration number and the added noise amplitude of the complete ensemble empirical mode decomposition with adaptive noise; The complexity-reduced historical data is decomposed into intrinsic mode function components and residual components by the complete ensemble empirical mode decomposition with adaptive noise; The zero-crossing rate and sample entropy are obtained from the intrinsic mode function components.

4. The load prediction method based on big data according to claim 1, characterized in that, The zero-crossing rate and sample entropy are used to determine the multi-frequency component of the load data, which comprises: The zero-crossing rate and sample entropy are used to determine the high-frequency component, the medium-frequency component and the low-frequency component of the load data; Wherein, the high-frequency component contains the influence of noise and sudden events, the medium-frequency component contains daily or weekly periodicity, and the low-frequency component contains long-term trend and seasonality.

5. The load prediction method based on big data according to claim 1, characterized in that, The prediction performance, model complexity and confidence of each prediction module in the multi-module weighted combination analysis model are determined according to the improved grey wolf algorithm, and the weight of the load prediction result of each prediction module is determined, which comprises: Assigning the initial weight of each prediction module in the multi-module weighted combination analysis model; Obtaining a prediction performance index, a model complexity and a confidence of each prediction gene module; An iterative update of the improved grey wolf algorithm is used in combination with a fitness function including the prediction performance index, the model complexity and the confidence to complete weight optimization of output results of each prediction gene module; The prediction gene module includes a wavelet transform and a convolutional neural network model, a recurrent neural network model and a random forest model.

6. The load forecasting method based on big data according to claim 5, characterized in that, The iterative update of the improved grey wolf algorithm is used in combination with the fitness function including the prediction performance index, the model complexity and the confidence to complete the weight optimization of the output results of each prediction gene module, including: Parameters of the improved grey wolf algorithm are configured, and a fitness value of each group of weight vectors is calculated according to the prediction performance index, the model complexity and the confidence of each prediction gene module and a corresponding weight vector; According to a preset update rule, positions of the wolf group are iteratively updated to find an optimal weight vector, and the weight vector with the highest fitness value is selected as the optimal weight.

7. The load prediction method based on big data according to claim 5, characterized in that, The fitness function is: Fitness = a prediction performance + β model complexity penalty + γ confidence adjustment fitness; In the formula, a, β and γ respectively represent weight vectors corresponding to the prediction performance, the model complexity penalty and the confidence adjustment fitness. 8.The load prediction method based on big data according to claim 5, wherein, The prediction performance index of each prediction gene module is obtained, including: A mean square error is used to calculate the prediction performance index of each prediction gene module on a test set.

9. A big data based load forecasting system for implementing the big data based load forecasting method of any one of claims 1-8, characterized in that, Including: A data preprocessing module is configured to perform nonlinear transformation on historical load data based on a least squares support vector machine, and to define a Lagrange function form to obtain a vector regression function expression to reduce the complexity of the historical load data; A data input matching module is configured to obtain a zero-crossing rate and a sample entropy by decomposing the historical load data with reduced complexity, and to determine multiple frequency components of the load data using the zero-crossing rate and the sample entropy; A model training module is configured to train a multi-gene weighted combination analysis model by using a preset hybrid algorithm in combination with the multiple frequency components of the load data, including: A wavelet transform and convolutional neural network model is trained by using a wavelet transform and a convolutional neural network in combination with high-frequency components to perform high-frequency component prediction; A recurrent neural network model is trained by using empirical mode decomposition and a recurrent neural network in combination with medium-frequency components to perform medium-frequency component prediction; a random forest model is trained by using Fourier transform and a random forest in combination with low-frequency components to perform low-frequency component prediction; and the wavelet transform and convolutional neural network model, the recurrent neural network model and the random forest model are combined to obtain the multi-gene weighted combination analysis model; A weight combination module is configured to determine weights of load prediction results of each prediction gene module based on prediction performance, model complexity and confidence of each prediction gene module in the multi-gene weighted combination analysis model and according to the improved grey wolf algorithm, and to obtain a final load prediction result by weighting and combining the load prediction results of each prediction gene module.

Images