Load prediction method and system based on improved CEEMDAN combination model

Abstract

The application discloses a load prediction method and system based on an improved CEEMDAN combination model, which decomposes load data into multiple basic modal components and a residual component by using the improved CEEMDAN model. The improved CEEMDAN model can effectively capture the nonlinear characteristics and time sequence information of the load time series data, and improve the accuracy of short-term load prediction. In addition, the multiple basic modal components and the residual component are respectively predicted by using a volatility weighted Lasso prediction model. The volatility weighted Lasso prediction model can dynamically adjust the shrinkage degree of the coefficient, and can show better robustness in the face of data volatility, noise and outliers. After obtaining the prediction results of each component, the final load prediction result is obtained by weighted summation, so that a more accurate and reliable model occupies a larger proportion in comprehensive prediction, and the overall accuracy of the prediction result is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power load prediction, in particular, to a load prediction method and system based on an improved CEEMDAN combination model, an electronic device and a computer readable storage medium. BACKGROUND

[0002] The current power system is facing increasing load demand, which is mainly affected by population growth, accelerated industrialization and improved electrification level, etc., resulting in increased volatility of load, especially in peak periods, higher requirements for the stability and reliability of the power grid. In addition, with the integration of renewable energy and the increase of distributed generation resources, the load characteristics of the power system become more complex and variable. Therefore, accurate short-term load forecasting is crucial for the operation and management of the power system, which helps to optimize power generation planning, improve energy utilization efficiency, reduce operating costs, and ensure the stability of power supply.

[0003] Most of the current load forecasting methods use quantitative models, such as time series methods, machine learning methods, deep learning methods, etc. Among them, the most representative time series method is the difference autoregressive moving average model, abbreviated as ARIMA model. The ARIMA model is a linear data-based model that cannot capture nonlinear data information. However, in reality, time series data often exhibit more nonlinearity and contain complex noise. Thus, the model order and reserved variables obtained by the linear model are not optimal, resulting in low prediction accuracy. Although machine learning methods have good nonlinear data fitting ability and can effectively handle nonlinear problems, they lack the ability to capture time series information and cannot fully utilize the time series information in the sequence. Even sometimes, human intervention is required to extract features. Deep learning methods generally use LSTM and GRU models. Compared with traditional RNN models, LSTM can solve the problems of gradient disappearance and gradient explosion caused by long-term dependence, and has stronger time series adaptation. However, it has a large number of network parameters, slow training speed, and limited network memory capacity when the input sequence is too long, which may lead to sequence loss and affect the accuracy of the model. GRU model is an optimized network based on LSTM, which has a simpler structure and reduces the training time of the model to some extent. However, the RNN network cannot be operated in parallel, which greatly wastes computing resources and makes it difficult to handle large amounts of power load data. Moreover, existing load forecasting methods usually use a single model for load forecasting, which has insufficient prediction accuracy, resulting in poor accuracy of short-term load forecasting. SUMMARY

[0004] The application provides a load prediction method and system based on an improved CEEMDAN combination model, an electronic device and a computer readable storage medium.

[0005] According to one aspect of the application, a load prediction method based on an improved CEEMDAN combination model is provided, comprising the following steps:

[0006] The load data is decomposed into a plurality of basic modal components and a residual component by using the improved CEEMDAN model.

[0007] The plurality of basic modal components and the residual component are respectively predicted by using a volatility weighted Lasso prediction model to obtain the prediction results of each component.

[0008] The prediction results of each component are weighted and summed to obtain the final load prediction result.

[0009] Further, the process of decomposing the load data into a plurality of basic modal components and a residual component by using the improved CEEMDAN model comprises the following steps:

[0010] A first set of white noise is added to the load time series to obtain a new sequence, and the new sequence is decomposed by EMD to obtain a first residual component and a first basic modal component.

[0011] A second set of white noise is added to the first residual component and EMD is performed again to obtain a second residual component and a second basic modal component.

[0012] The above steps are iterated continuously, and whether the iteration termination condition is met in the iteration process is judged, and if the iteration termination condition is met, the plurality of basic modal components and the residual component are output.

[0013] Further, the sample entropy is used as the termination condition in the iteration process, and the specific judgment process is as follows:

[0014] The residual component obtained by iteration is obtained, and the embedding dimension, tolerance and sample entropy threshold are set.

[0015] The obtained residual component is embedded and reconstructed according to the set embedding dimension to obtain a plurality of embedding vectors.

[0016] The similarity between any two embedding vectors is calculated, and the number of embedding vector pairs satisfying the tolerance condition is counted.

[0017] The embedding dimension is added by one, and the obtained residual component is embedded and reconstructed using the new embedding dimension. The number of new embedding vector pairs satisfying the tolerance condition of similarity is also counted.

[0018] The sample entropy is calculated based on the number of two embedding vector pairs. The calculated sample entropy is compared with the set sample entropy threshold. If the calculated sample entropy is less than or equal to the sample entropy threshold, it is determined that the termination condition is met. If the calculated sample entropy is greater than the sample entropy threshold, it is determined that the termination condition is not met.

[0019] Further, the process of predicting a plurality of basic modal components and a residual component using the volatility weighted Lasso prediction model to obtain the prediction result of each component includes the following contents:

[0020] A Lasso regression model is constructed for each component, and the order of each Lasso regression model is determined.

[0021] The data of each component is feature extracted using the Lasso regression model, and the feature reconstruction is performed according to the determined model order to obtain the feature reconstruction matrix of each component.

[0022] For each feature reconstruction matrix, the volatility of each feature is estimated by performing volatility estimation on each feature of the matrix, and the volatility of each feature is added to the penalty term of the objective function of the Lasso regression model, so as to construct a volatility weighted Lasso model and solve the objective function thereof, thereby obtaining the prediction result of each component.

[0023] Further, the expression of the objective function is:

[0024]

[0025] Wherein, y i represents the observation value of the i th sample, X i represents the feature vector of the i th sample, m represents the sample number, β represents the coefficient vector, λ represents the regularization parameter, p represents the feature number, β j represents the coefficient of the j th feature, represents the volatility of the j th feature.

[0026] Further, the process of determining the order of each Lasso regression model is:

[0027] Based on the data of each component, leave-one-out test is performed on the Lasso regression model in ascending order from low order to high order, the mean square errors of the Lasso regression models of adjacent two orders are compared, if the mean square error value of the Lasso regression model of the previous order is greater than the mean square error value of the Lasso regression model of the next order, the order is continued to expand, if the mean square error value of the Lasso regression model of the previous order is less than or equal to the mean square error value of the Lasso regression model of the next order, the order expansion is terminated, and the order of the previous order is taken as the model order.

[0028] Further, the final load prediction result is calculated based on the following formula:

[0029]

[0030] Wherein, lotal weight represents the load prediction value calculated by weighted summation, wm j represents the weight of the jth fluctuation weighted Lasso model, ymj represents the load prediction value of the jth fluctuation weighted Lasso model, em j represents the average error of the jth fluctuation weighted Lasso model, and k represents the number of fluctuation weighted Lasso models.

[0031] In addition, the present application also provides a load prediction system based on the improved CEEMDAN combined model, comprising:

[0032] A data decomposition module is configured to decompose the load data into a plurality of basic modal components and a residual component by using the improved CEEMDAN model.

[0033] A component prediction module is configured to predict the plurality of basic modal components and the residual component by using the fluctuation weighted Lasso prediction model, so as to obtain the prediction results of each component.

[0034] A weighted summation module is configured to perform weighted summation on the prediction results of each component, so as to obtain the final load prediction result.

[0035] In addition, the present application also provides an electronic device comprising a processor and a memory, wherein the memory stores a computer program, and the processor is configured to execute the steps of the method described above by calling the computer program stored in the memory.

[0036] In addition, the present application also provides a computer readable storage medium for storing a computer program for load prediction based on the improved CEEMDAN combined model, wherein the computer program executes the steps of the method described above when running on a computer.

[0037] The present application has the following advantages:

[0038] The load prediction method based on the improved CEEMDAN combined model of the present application first decomposes the load data into a plurality of basic modal components and a residual component by using the improved CEEMDAN model. The improved CEEMDAN model can effectively capture the nonlinear characteristics and time sequence information of the load time series data, thereby improving the accuracy of short-term load prediction. Moreover, the plurality of basic modal components and the residual component are respectively predicted by using the volatility weighted Lasso prediction model. The volatility weighted Lasso prediction model can dynamically adjust the shrinkage degree of the coefficient, and can exhibit better robustness in the face of data volatility, noise and outliers. After obtaining the prediction results of each component, the final load prediction result is obtained by weighted summation, so as to ensure that the more accurate and reliable model occupies a larger proportion in the comprehensive prediction, thereby improving the overall accuracy of the prediction result.

[0039] In addition, the load prediction system based on the improved CEEMDAN combined model of the present application also has the above advantages.

[0040] In addition to the purposes, features and advantages described above, the present application has other purposes, features and advantages. The present application will be further described below with reference to the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0041] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application, illustrate the preferred embodiment of the application and assist in the explanation of the application. In the drawings:

[0042] Figure 1 is a flowchart of the load prediction method based on the improved CEEMDAN combined model of the preferred embodiment of the present application.

[0043] Figure 2 is a flowchart of the load prediction method based on the improved CEEMDAN combined model of the preferred embodiment of the present application. Figure 1 is a sub-flowchart of step S1 in the load prediction method based on the improved CEEMDAN combined model of the preferred embodiment of the present application.

[0044] Figure 3 is a sub-flowchart of step S13 in the load prediction method based on the improved CEEMDAN combined model of the preferred embodiment of the present application. Figure 2 is a sub-flowchart of step S13 in the load prediction method based on the improved CEEMDAN combined model of the preferred embodiment of the present application.

[0045] Figure 4 is a sub-flowchart of step S2 in the load prediction method based on the improved CEEMDAN combined model of the preferred embodiment of the present application. Figure 1 is a sub-flowchart of step S2 in the load prediction method based on the improved CEEMDAN combined model of the preferred embodiment of the present application.

[0046] Figure 5 is a module structure schematic diagram of the load prediction system based on the improved CEEMDAN combined model of another embodiment of the present application. DETAILED DESCRIPTION

[0047] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the drawings and in combination with the embodiments.

[0048] With reference to Figure 1 , the preferred embodiments of the present application provide a load prediction method based on an improved CEEMDAN combination model, including the following contents:

[0049] Step S1: decompose the load data into a plurality of basic modal components and a residual component by using an improved CEEMDAN model;

[0050] Step S2: respectively predict the plurality of basic modal components and the residual component by using a volatility weighted Lasso prediction model to obtain the prediction results of each component;

[0051] Step S3: weighted sum the prediction results of each component to obtain the final load prediction result.

[0052] It can be understood that the load prediction method based on the improved CEEMDAN combination model of the present embodiment first decomposes the load data into a plurality of basic modal components and a residual component by using an improved CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) model. The improved CEEMDAN model can effectively capture the nonlinear characteristics and time series information of the load time series data, thereby improving the accuracy of short-term load prediction. Moreover, the plurality of basic modal components and the residual component are respectively predicted by using a volatility weighted Lasso prediction model. The volatility weighted Lasso prediction model can dynamically adjust the shrinkage degree of the coefficient, and can exhibit better robustness in the face of data volatility, noise and outliers. After obtaining the prediction results of each component, the final load prediction result is obtained by weighted sum, so as to ensure that the more accurate and reliable model occupies a larger proportion in the comprehensive prediction, thereby improving the overall accuracy of the prediction result.

[0053] It can be understood that, as Figure 2 shown, in the step S1, the process of decomposing the load data into a plurality of basic modal components and a residual component by using the improved CEEMDAN model includes the following contents:

[0054] Step S11: add a first group of white noise to the load time series to obtain a new sequence, and perform EMD decomposition on the new sequence to obtain a first residual component and a first basic modal component;

[0055] Step S12: adding a second set of white noise in the first residual component and performing EMD decomposition again to obtain a second residual component and a second intrinsic mode component;

[0056] Step S13: continuously iterating the above steps and judging whether the iteration termination condition is met in the iteration process, and outputting a plurality of intrinsic mode components and a residual component if the iteration termination condition is met.

[0057] Specifically, a first set of white noise is added in the obtained load time series x to obtain a new load time series, which can be represented as: x (u) = x + β0E1[w (u) ], wherein x (u) represents the new load time series, x represents the original load time series, i.e. the original signal, β0represents the expected signal-to-noise ratio of the first decomposition signal, E1[w (u) ] represents the first set of white noise, w (u) represents the u-th added white noise, u = 1, 2, …, U, and U represents the number of white noises. Then, the new sequence x (u) is subjected to EMD decomposition to obtain a first residual component and a first intrinsic mode component, wherein the first residual component can be represented as: The first intrinsic mode component can be represented as: M[] represents calculating the local mean M times. Then, a second set of white noise is added in the obtained first residual component and EMD decomposition is performed again to obtain a second residual component and a second intrinsic mode component, wherein the second residual component can be represented as: The second intrinsic mode component can be represented as: E2[w (u) ] represents the second set of white noise. Then, the above steps are continuously iterated until the iteration termination condition is met, and a plurality of intrinsic mode components and a residual component are outputted.

[0058] It can be understood that the EMD decomposition can decompose the signal according to the time scale characteristics of the signal itself without setting any base function in advance, and the EMD decomposes the time sequence into several IMF components (i.e. intrinsic mode components) and a residual, each component is called an intrinsic mode function, the IMF component represents the fluctuation component in the original signal, and the residual component represents the basic trend of the original signal. And EEMD (Ensemble Empirical Mode Decomposition, Ensemble Empirical Mode Decomposition) decomposition is to add noise auxiliary data decomposition on the basis of EMD decomposition method, its principle is that when the signal is mapped to the corresponding scale by adding uniformly distributed white noise, the added noise will be eliminated when the overall mean is calculated using a sufficient test set, thereby effectively suppressing the generation of mode mixing. And CEEMDAN is to add a limited adaptive white noise in EEMD, which overcomes the problem of incomplete EEMD and reconstruction error after adding white noise, and adaptive noise complete empirical mode decomposition not only eliminates the mode mixing and noise residual of EMD and EEMD to a great extent. However, in the existing CEEMDAN method, in the process of decomposing the load data into a series of intrinsic mode components (IMFs) through EEMD, some residual noise may be contained in each IMF, which affects the accuracy of subsequent analysis, especially in signal reconstruction or feature extraction; At the same time, some modes that do not meet the definition of IMF may also be generated, which causes confusion in the physical interpretation and analysis of the signal, and also affects the accuracy of subsequent analysis; In addition, there is still a problem of mode mixing in the high frequency component of the signal. The improved CEEMDAN model of the present application controls the signal-to-noise ratio of each decomposition by adding different levels of white noise in each iteration, so that the signal obtained in each iteration is closer to its internal oscillation structure, thereby improving the accuracy of signal decomposition. This iterative method can gradually separate different time scale oscillation modes from the original signal, realize accurate decomposition, obtain clearer and more effective IMF components, effectively solve the problems of mode mixing and noise residue in the high frequency component of the existing CEEMDAN, improve the quality of signal decomposition, and thereby improve the load prediction accuracy.

[0059] Optionally, in the step S13, the sample entropy is used as a termination condition in the iteration process, as shown in the following formula: Figure 3 As shown in the following formula, the specific judgment process is as follows:

[0060] Step S131: obtaining the residual component obtained by iteration, and setting embedding dimension, tolerance and sample entropy threshold;

[0061] Step S132: embedding reconstruction is performed on the obtained residual component according to the set embedding dimension, to obtain a plurality of embedding vectors;

[0062] Step S133: Calculate the similarity between any two embedding vectors, and count the number of embedding vector pairs that satisfy the tolerance condition of similarity;

[0063] Step S134: Add one to the embedding dimension, and use the new embedding dimension to reconstruct the obtained residual component, and similarly count the number of new embedding vector pairs that satisfy the tolerance condition of similarity;

[0064] Step S135: Calculate the sample entropy based on the number of two embedding vector pairs, and compare the calculated sample entropy with the set sample entropy threshold value. If the calculated sample entropy is less than or equal to the sample entropy threshold value, it is determined that the termination condition is satisfied. If the calculated sample entropy is greater than the sample entropy threshold value, it is determined that the termination condition is not satisfied.

[0065] Specifically, assuming that the iteratively obtained residual components are x = [1, 2, 4, 5, 7, 9.2, 13], and the embedding dimension CrWd, the tolerance YbRrd, and the sample entropy threshold YbFzd are set as 2, 0.5, and 0.5 respectively, where the embedding dimension CrWd determines the number of points considered in the time series, the tolerance YbRrd determines the threshold value at which two subsequences are considered similar, and the sample entropy threshold YbFzd determines whether the residual components are to continue to be decomposed. Then, the residual components x are reconstructed by embedding with the embedding dimension = 2, to generate a plurality of embedding vectors, which can be represented as Y(1) = [1, 2], Y(2) = [2, 4], Y(3) = [4, 5], Y(4) = [5, 7], Y(5) = [7, 9.2], and Y(6) = [9.2, 13]. Then, the similarity between any two embedding vectors is calculated, for example, the distance between two vectors is calculated by using the Euclidean distance to evaluate the similarity, and when the distance value between two vectors is less than or equal to the tolerance YbRrd, it is determined that the two vectors are similar. After calculation, the similarity results are obtained as follows: for Y(1), only Y(2) is similar within YbRrd; for Y(2), only Y(1) and Y(3) are similar within YbRrd; for Y(3), only Y(2) and Y(4) are similar within YbRrd; for Y(4), only Y(3) and Y(5) are similar within YbRrd; for Y(5), only Y(4) and Y(6) are similar within YbRrd; and for Y(6), there is no other vector similar within YbRrd. Therefore, there are 5 pairs of vectors that satisfy the tolerance condition, and thus the number of embedding vector pairs B(CrWd, YbRrd) that satisfy the tolerance condition is 5. Next, the embedding dimension is increased by one, that is, the embedding dimension is set to 3, and then the embedding reconstruction is performed, to obtain a plurality of embedding vectors, which are Y(1) = [1, 2, 4], Y(2) = [2, 4, 5], Y(3) = [4, 5, 7], Y(4) = [5, 7, 9.2], and Y(5) = [7, 7.9, 13]. Similarly, after the similarity is judged, it is obtained that: for Y(1), there is no other vector similar within YbRrd; for Y(2), only Y(3) is similar within YbRrd; for Y(3), Y(2) and Y(4) are similar within YbRrd; for Y(4), Y(3) and Y(5) are similar within YbRrd; and for Y(5), there is no other vector similar within YbRrd. Therefore, there are 3 pairs of vectors that satisfy the tolerance condition, that is, the new number of embedding vector pairs B(CrWd+1, YbRrd) is 3. Then, the sample entropy is calculated based on the following formula: SampEn = -ln((B(CrWd+1, YbRrd)) / (B(CrWd, YbRrd))), where SampEn represents the calculated sample entropy, and it can be calculated that SampEn = -ln(3 / 5) = -ln(0.6) ≈ 0.5108 at this time.Then, the calculated sample entropy SampEn is compared with a set sample entropy threshold YbFzd, when the calculated sample entropy SampEn is less than or equal to YbFzd, the residual sequence will not be decomposed, when the calculated sample entropy SampEn is greater than YbFzd, the residual sequence will continue to be decomposed. For example, the calculated sample entropy SampEn = 0.5108 is greater than the sample entropy threshold 0.5, and the residual component x needs to be decomposed.

[0066] It can be understood that in the existing CEEMDAN, the termination condition is that the algorithm is terminated until the residual component becomes a monotonic function or cannot be further decomposed into IMF, which may cause excessive iteration and unnecessary calculation amount, and the calculation efficiency is low. The sample entropy strategy is introduced as the termination condition in the present application, and the complexity of the residual component obtained by iteration is evaluated by sample entropy to determine whether to terminate iteration, when the complexity (i.e. the calculated sample entropy) of the residual component is lower than the set complexity tolerance (i.e. the sample entropy threshold), the iteration is terminated, the calculation amount is reduced, the error accumulation in the subsequent model is avoided, and the calculation efficiency is improved.

[0067] It can be understood that, as shown in Figure 4 In the step S2, the process of predicting a plurality of basic modal components and a residual component by using the volatility weighted Lasso prediction model to obtain the prediction results of each component includes the following contents:

[0068] Step S21: constructing a Lasso regression model for each component respectively, and determining the order of each Lasso regression model;

[0069] Step S22: performing feature extraction on the data of each component by using the Lasso regression model, and performing feature reconstruction according to the determined model order to obtain a feature reconstruction matrix of each component;

[0070] Step S23: for each feature reconstruction matrix, volatility estimation is performed on each feature of the matrix to obtain the volatility of each feature, and the volatility of each feature is added to the penalty term of the objective function of the Lasso regression model, so as to construct a volatility weighted Lasso model and solve the objective function thereof, thereby obtaining the prediction result of each component.

[0071] It can be understood that the purpose of the present application is to obtain a weighted combination of future load, when the future prediction value corresponding to each signal component is known, and the future value of each signal component is predicted by a volatility weighted Lasso (Least absolute shrinkage and selection operator, Least absolute shrinkage and selection operator regression) regression model, so it is necessary to reconstruct the data of each signal component, which is implemented in two parts, one is the order determination of the signal component data model, and the other is the reconstruction of the signal component sequence. Therefore, in the step S1, the process of determining the order of each Lasso regression model is:

[0072] Based on the data of each component, the leave-one-out method is tested by Lasso regression model from low order to high order, the mean square errors of the Lasso regression models of adjacent two orders are compared, if the mean square error value of the Lasso regression model of the previous order is greater than that of the Lasso regression model of the next order, the order is continued to expand, if the mean square error value of the Lasso regression model of the previous order is less than or equal to that of the Lasso regression model of the next order, the order expansion is terminated, and the previous order is taken as the model order.

[0073] Specifically, assuming that the data sequence of each component includes N samples, one dependent variable and g-1 independent variables, then the leave-one-out method is tested by Lasso regression model from low order to high order, and whether to expand the order is determined according to the minimum RMSE standard. Let RMSE lasso(n) be the mean square error of the Lasso regression model with order n, RMSE lasso(n+1) be the mean square error of the Lasso regression model with order n+1, then the judgment formula is: if The order of each Lasso regression model can be obtained by the above method of determining the order of the data of each signal component, which can be represented as: Jieshu={jieshu1,jieshu2,jieshu3,...,jieshu L}, wherein Jieshu represents the set of signal component model order determination, jieshu L represents the order of the Lasso regression model corresponding to the Lth signal component. The leave-one-out method belongs to the existing method, and the specific test process is not described here.

[0074] It can be understood that the existing model order determination method has two kinds, one is to select by experience, and the other is to set other parameters of the model first, and then optimize the lag order according to certain standards. Among them, the first method excessively depends on the level and experience of the researchers, and cannot objectively select the best lag stage; and the second method ignores an important problem: the influence of the lag order and other parameters on the model is mutual, if the lag order is selected by setting other parameters first, and then the parameter is optimized using the lag order value, it is likely that the parameter is only optimal under the lag order, but not globally optimal. In the present application, by optimizing the model order together with the Lasso model parameter, the effect of global optimization can be achieved.

[0075] It can be understood that in the step S22, from the definition of model order determination, it can be known that the current value of each signal component is affected by the previous several values, for example, the order of a certain component is w, which means that the values of the previous w days are used as the input of the Lasso regression model to predict the value of the current day. Assuming that the feature data sequence extracted by a certain Lasso regression model is Fet={fet1, fet2,.., fet N}, and the model order w determined in step S21 is combined to reconstruct the features, at this time the input sample is converted into a w-column matrix, the output sample is a 1-column matrix, the input and the input matrix are combined, the first w columns are features, and the last 1 column is the target, so that the feature reconstruction matrix is constructed, which can be represented as: Wherein, sample w represents the feature reconstruction matrix of the signal component with the order w, N is the number of samples, fet w represents the wth value corresponding to the feature. Similarly, based on the order set Jieshu obtained by the above method combined with step S21, the feature reconstruction matrix set of each signal component can be obtained, which can be represented as: Wherein, Chgjz represents the feature reconstruction matrix set of L signal components, represents the feature reconstruction matrix of the Lth signal component. It can be understood that by determining the order of the signal component data model and reconstructing the feature sequence, the future value of each signal component can be more accurately predicted.

[0076] It can be understood that in the step S23, for each feature reconstruction matrix, the volatility of each feature (i.e. each independent variable) of the matrix is estimated by using the GARCH(1,1) model, and the specific process is as follows:

[0077] First, set the model, and the calculation formula of volatility is as follows:

[0078] r t = μ + ε t

[0079] ε t = σ t z t

[0080]

[0081] where r t denotes the observation of the time series at time t, μ denotes the mean of the observations of the time series, ε t denotes the error term of the time series at time t, σ t denotes the conditional variance, z t denotes the volatility at time t, z denotes the independent and identically distributed random error term, which follows a standard normal distribution,

[0082] denotes the volatility, a0 denotes the constant term, which represents the baseline level of the volatility, a q and h d respectively denote the past volatility and the past volatility's impact on the current volatility, ε t-q denotes the past error term. Then the gradient descent method is used to estimate the parameters a0, a q and h d , and the estimated parameters are used to predict the volatility corresponding to each independent variable based on the following formula:

[0083]

[0084] Next, the volatility of each feature is added to the penalty term of the Lasso regression model objective function in the form of an inverse, where the expression of the objective function is as follows:

[0085] where y i denotes the observation of the i-th sample, X i denotes the feature vector of the i-th sample, m denotes the number of samples, β denotes the coefficient vector, λ denotes the regularization parameter, p denotes the number of features, β j denotes the coefficient of the j-th feature, and z denotes the volatility of the j-th feature. Then the existing coordinate descent method or least angle regression method is used to iteratively solve the objective function to obtain the adjusted coefficient vector β and the predicted value. It can be understood that the above process is performed for each matrix in the feature reconstruction matrix set obtained in step S22, so that the volatility weighted Lasso model and the predicted value corresponding to each signal component can be constructed.

[0086] It can be understood that, on the basis of the existing Lasso model, the volatility weighting is taken as the weight of the model, that is, the volatility of the feature is taken as a component of the penalty term of the objective function, so that the shrinkage degree of the coefficient in the model can be adjusted, the penalty on the coefficient is stricter when the data volatility is higher, and the penalty on the coefficient is relatively loose when the volatility is lower, so that the model is more robust when facing data volatility, and the sensitivity to data is maintained when the volatility is low, thereby improving the robustness of the model when facing data volatility, noise and outliers.

[0087] It can be understood that, in the step S3, the final load prediction result is calculated based on the following formula:

[0088]

[0089] wherein, lotal weight represents the load prediction value calculated by weighted summation, wm j represents the weight of the jth volatility weighted Lasso model, ym j represents the load prediction value of the jth volatility weighted Lasso model, em j represents the average error of the jth volatility weighted Lasso model, and k represents the number of volatility weighted Lasso models.

[0090] Specifically, for the multiple volatility weighted Lasso models and their prediction values constructed in step S2, the average error value of each model is first calculated, and the reciprocal of each average error value is calculated, wherein the average error value calculation method belongs to the existing method, and will not be described here. Then, the weights of each volatility weighted Lasso model are calculated by adding and normalizing all the reciprocals of the average error values, which can be represented as: Finally, the final load prediction value is obtained by weighted summation based on the weight and prediction value of each volatility weighted Lasso model, and the calculation formula is:

[0091] It can be understood that the present application can automatically adjust the weight of the model according to the performance of the volatility weighted Lasso model, so that the more accurate and reliable model occupies a larger proportion in the comprehensive prediction, and the overall accuracy of the prediction result is improved.

[0092] In addition, as Figure 5 shown, another embodiment of the present application also provides a load prediction system based on an improved CEEMDAN combined model, preferably using the load prediction method as described above, which comprises:

[0093] A data decomposition module is configured to decompose the load data into a plurality of basic modal components and a residual component by using the improved CEEMDAN model.

[0094] a component prediction module configured to predict the plurality of basic modal components and the residual component respectively by using a volatility weighted Lasso prediction model, and obtain prediction results of the components;

[0095] a weighted summation module configured to perform weighted summation on the prediction results of the components, and obtain a final load prediction result.

[0096] It can be understood that the load prediction system based on the improved CEEMDAN combined model in the embodiment first decomposes the load data into a plurality of basic modal components and a residual component by using the improved CEEMDAN model. The improved CEEMDAN model can effectively capture the nonlinear characteristics and time sequence information of the load time series data, and improve the accuracy of short-term load prediction. Furthermore, the plurality of basic modal components and the residual component are predicted respectively by using the volatility weighted Lasso prediction model. The volatility weighted Lasso prediction model can dynamically adjust the shrinkage degree of the coefficient, and can show better robustness in the face of data volatility, noise and outliers. After obtaining the prediction results of the components, the final load prediction result is obtained by performing weighted summation, so as to ensure that the more accurate and reliable model occupies a larger proportion in the comprehensive prediction, and the overall accuracy of the prediction result is improved.

[0097] In addition, another embodiment of the present application further provides an electronic device comprising a processor and a memory, wherein the memory stores a computer program, and the processor is configured to execute the steps of the method described above by calling the computer program stored in the memory.

[0098] In addition, another embodiment of the present application further provides a computer readable storage medium for storing a computer program for load prediction based on an improved CEEMDAN combined model, wherein the computer program executes the steps of the method described above when running on a computer.

[0099] Generally, computer readable storage media includes any media that can be accessed by a computer. By way of example, and not limitation, computer readable storage media can comprise storage media such as RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other storage medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer. Also, any connection is properly termed a computer readable medium. For example, if the software is transmitted from a website, server, or other remote source using a coaxial cable, fiber optic cable, or twisted pair, then the coaxial cable, fiber optic cable, or twisted pair are included in the definition of medium. Disk and disc, as used herein, includes compact disc (CD), laser disc, optical disc, Blu-ray® disc, and floppy disk used to store software.

[0100] Those skilled in the art will appreciate that embodiments of the present application can be devised for a method, a system, or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer readable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable program code thereon for use by or in connection with an instruction execution system. Program code embodied on one or more computer readable storage media can be transmitted using any

[0101] The computer program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other devices to cause a series of operational steps to be performed on the computer, other programmable apparatus or other devices to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks. Figure 1 The flowchart and / or block diagram in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present application. In this regard, each block in the flowchart and / or block diagrams can represent a module, segment, or portion of instructions, which comprises one or more Figure 1 The functions indicated in each block or blocks of the flowchart and / or block diagram can be combined with one or more other blocks in the flowchart and / or block diagram. Furthermore, the functions indicated in each block or blocks of the flowchart and / or block diagram can be implemented in hardware, software, or a combination of hardware and software.

[0102] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0103] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0104] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0105] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

[0106] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A load forecasting method based on an improved CEEMDAN combined model, characterized in that, Includes the following: The load data is decomposed into multiple basic modal components and a residual component using an improved CEEMDAN model. The volatility-weighted Lasso prediction model was used to predict multiple fundamental modal components and one residual component, and the prediction results of each component were obtained. The prediction results of each component are weighted and summed to obtain the final load prediction result; The process of decomposing load data into multiple fundamental modal components and a residual component using the improved CEEMDAN model includes the following: The first set of white noise is added to the load time series to obtain a new series, and the new series is decomposed by EMD to obtain the first residual component and the first fundamental mode component. A second set of white noise is added to the first residual component and EMD decomposition is performed again to obtain the second residual component and the second fundamental mode component. The above steps are iterated continuously, and the iteration termination condition is checked during the iteration process. If the iteration termination condition is met, multiple basic modal components and one residual component are output. The process of using a volatility-weighted Lasso prediction model to predict multiple fundamental modal components and one residual component, and obtaining the prediction results for each component, includes the following: For each component, construct a Lasso regression model and determine the order of each Lasso regression model; The Lasso regression model is used to extract features from the data of each component, and the features are reconstructed according to the determined model order to obtain the feature reconstruction matrix of each component. For each feature, a reconstructed matrix is ​​obtained by estimating the volatility of each feature. The volatility of each feature is then added to the penalty term of the objective function of the Lasso regression model, thereby constructing a volatility-weighted Lasso model and solving its objective function to obtain the prediction result for each component. The expression for the objective function is: Among them, y i Let X represent the observation value of the i-th sample. i Let m represent the number of samples, β represent the coefficient vector, λ represent the regularization parameter, and p represent the number of features. j The coefficient of the j-th feature is represented by the coefficient of the j-th feature. This represents the volatility of the j-th feature.

2. The load forecasting method based on the improved CEEMDAN combined model as described in claim 1, characterized in that, The sample entropy is used as the termination condition during the iteration process. The specific judgment process is as follows: Obtain the residual components obtained from the iteration, and set the embedding dimension, tolerance, and sample entropy threshold; The obtained residual components are reconstructed by embedding according to the set embedding dimension to obtain multiple embedding vectors; Calculate the similarity between any two embedding vectors and count the number of embedding vector pairs whose similarity satisfies the tolerance condition; The embedding dimension is incremented by one, and the new embedding dimension is used to reconstruct the embedding of the obtained residual components. Similarly, the number of new embedding vector pairs whose similarity satisfies the tolerance condition is counted. The sample entropy is calculated based on the number of two embedded vector pairs. The calculated sample entropy is then compared with a predefined sample entropy threshold. If the calculated sample entropy is less than or equal to the sample entropy threshold, the termination condition is satisfied. If the calculated sample entropy is greater than the sample entropy threshold, the termination condition is not satisfied.

3. The load forecasting method based on the improved CEEMDAN combined model as described in claim 1, characterized in that, The process of determining the order of each Lasso regression model is as follows: Based on the data for each component, the Lasso regression model is tested with leave-one-out method incrementally from low to high order. The mean squared error of the Lasso regression model of two adjacent orders is compared. If the mean squared error of the Lasso regression model of the previous order is greater than that of the Lasso regression model of the next order, the order is increased. If the mean squared error of the Lasso regression model of the previous order is less than or equal to that of the Lasso regression model of the next order, the order increase is terminated, and the previous order is taken as the model order.

4. The load forecasting method based on the improved CEEMDAN combined model as described in claim 1, characterized in that, The final load forecast result is calculated based on the following formula: Among them, lotal weight wm represents the load forecast value obtained by weighted summation. j Let ym represent the weights of the j-th volatility-weighted Lasso model. j Let em represent the load forecast value of the j-th volatility-weighted Lasso model. j Let represent the average error of the j-th volatility-weighted Lasso model, and k represent the number of volatility-weighted Lasso models.

5. A load forecasting system based on an improved CEEMDAN combined model, employing the load forecasting method based on the improved CEEMDAN combined model as described in any one of claims 1 to 4, characterized in that, include: The data decomposition module is used to decompose load data into multiple basic modal components and a residual component using an improved CEEMDAN model. The component prediction module is used to predict multiple basic modal components and one residual component using the volatility-weighted Lasso prediction model, and obtain the prediction results for each component. The weighted summation module is used to sum the prediction results of each component in a weighted manner to obtain the final load prediction result.

6. An electronic device, characterized in that, The method includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the steps of the method as described in any one of claims 1 to 4 by calling the computer program stored in the memory.

7. A computer-readable storage medium for storing a computer program for load forecasting based on an improved CEEMDAN combined model, characterized in that, The computer program, when run on a computer, performs the steps of the method as described in any one of claims 1 to 4.

Images