Power load prediction method and device based on hybrid expert decomposition

Abstract

The invention relates to a power load prediction method and device based on hybrid expert decomposition. According to the method, an original data set is obtained and preprocessed, wherein the original data set comprises power load historical time sequence data and weather data; performing correlation analysis on the preprocessed power load historical time sequence data and weather data, and selecting corresponding power load historical time sequence data under a positive correlation dimension; and decomposing the selected historical time sequence data of the power load into a season term and a trend term by using a hybrid expert decomposition and prediction model, respectively performing prediction, and splicing different prediction results to obtain a final power load prediction result. Compared with the prior art, the method has the advantages of improving the accuracy of power load prediction and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of electric load forecasting, and in particular to an electric load forecasting method and device based on hybrid expert decomposition. Background Art

[0002] The prediction of power load fluctuations in power networks has always been the core of research in this field. With the advancement of information collection and prediction algorithm technology, power load estimation methods are also evolving, covering a variety of methods such as traditional statistical analysis, machine intelligence, and deep neural networks. However, the growth in data volume and the increasing diversity of load patterns have brought certain limitations to existing prediction methods. For example, traditional statistical learning methods are difficult to accurately handle nonlinear and high-dimensional features, and the prediction accuracy of neural network-based prediction models may decrease when processing new data. Therefore, in the face of the complex and changeable load patterns of power systems and the growth in data volume, how to further improve the accuracy of power load prediction has become a problem that needs to be solved in this field. Summary of the invention

[0003] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a power load forecasting method and device based on hybrid expert decomposition, which can effectively extract potential load information when facing the complex and changeable load patterns of the power system and the growth of data volume, thereby improving the accuracy of power load forecasting.

[0004] The purpose of the present invention can be achieved by the following technical solutions:

[0005] According to a first aspect of the present invention, there is provided a method for power load forecasting based on hybrid expert decomposition, comprising the following steps: S1, obtaining and preprocessing an original data set, wherein the original data set comprises power load historical time series data and weather data; S2, performing correlation analysis on the preprocessed power load historical time series data and weather data, and selecting the corresponding power load historical time series data under the positive correlation dimension; S3, decomposing the selected power load historical time series data into seasonal items and trend items by using a hybrid expert decomposition and forecasting model, and forecasting them separately, and obtaining the final power load forecasting result after splicing different forecasting results.

[0006] As an optimal technical solution, the hybrid expert decomposition and prediction model includes a decomposition block, a seasonal item prediction block, a trend item prediction block and a splicing block, and S3 specifically includes: the decomposition block uses an expert hybrid strategy to decompose the selected power load historical time series data into seasonal items and trend items; the trend item prediction block uses a damped trend method to predict the trend item to obtain a trend item prediction result; the seasonal item prediction block uses a multi-scale extended inner layer convolution to capture local features and global correlations and make predictions to obtain a seasonal item prediction result; the splicing block splices the trend item prediction result with the seasonal item prediction result to obtain a final power load forecast result.

[0007] As a preferred technical solution, the process of predicting the trend item using the damped trend method is expressed as:

[0008] S t =αX t +(1-α)(S t-1 +φb t-1 )

[0009] b t =β(S t -S t-1 )+(1-β)φb t-1

[0010] In the formula, S t represents the estimated value of the historical time series of power load, b t represents the slope estimate of the historical time series of power load, the subscripts t and t-1 represent the t time step and t-1 time step respectively, α is the level of the smoothing parameter, β is the smoothing parameter of the trend, φ is the supplementary parameter, and X t For trend items.

[0011] As a preferred technical solution, the process of using multi-scale extended inner convolution to capture local features and global correlation and make predictions is expressed as:

[0012]

[0013] Y s,l =Function(Y s,l-1 ),l∈{1,2,…,N}

[0014] Y s =Truncate(Projection(Y s,N )),

[0015] In the formula, X zero ∈R O×d represents a placeholder filled with zeros, Represents X sThe embedding layer, Y s,l ∈R (I+O)×D represents the output of the lth multi-scale expansion layer within the convolutional layer, Y s ∈R O×d Represents the linear function projection Y s,N ∈R (I +O)×D Final forecast and truncation operations for the post-seasonal portion, represents the predicted value of the first layer, N represents the number of embedding layers, I, O, and D represent the lth multi-scale convolutional layer, 0 value, and latitude, respectively.

[0016] As a preferred technical solution, the embedding process is expressed as:

[0017]

[0018] Where TFE represents temporal feature encoding, CVE represents static covariate information encoding, and VALUE represents value embedding.

[0019] As a preferred technical solution, the preprocessing process specifically includes missing value processing and feature normalization.

[0020] As a preferred technical solution, S2 specifically includes: obtaining corresponding weather dimension information based on the preprocessed power load historical time series data; performing correlation analysis on the preprocessed power load historical time series data and the corresponding weather dimension information, and calculating the correlation coefficient between each weather dimension information and the corresponding power load historical time series data; based on the correlation coefficient, selecting the power load historical time series data corresponding to the dimension that is positively correlated with the power load historical time series data.

[0021] As a preferred technical solution, the correlation analysis is performed based on a random forest method.

[0022] According to a second aspect of the present invention, there is provided an electric load forecasting device based on hybrid expert decomposition, comprising: a data acquisition and preprocessing module, for acquiring and preprocessing an original data set, wherein the original data set comprises electric load historical time series data and weather data; a data screening module, for performing correlation analysis on the preprocessed electric load historical time series data and weather data, and selecting the electric load historical time series data corresponding to the positive correlation dimension; and a decomposition and prediction module, for decomposing the selected electric load historical time series data into seasonal items and trend items by using a hybrid expert decomposition and prediction model, and performing predictions respectively, and obtaining a final electric load forecasting result after splicing different prediction results.

[0023] As a preferred technical solution, the decomposition and prediction module includes: a decomposition unit, which is used to decompose the selected power load historical time series data into seasonal items and trend items using an expert hybrid strategy; a trend item prediction unit, which is used to predict the trend item using a damped trend method to obtain a trend item prediction result; a seasonal item prediction unit, which is used to capture local features and global correlations using a multi-scale extended inner layer convolution and perform predictions to obtain a seasonal item prediction result; and a splicing unit, which is used to splice the trend item prediction result with the seasonal item prediction result to obtain a final power load prediction result.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] 1. The present invention uses a hybrid expert decomposition and prediction model to split the historical time series data of power load into two components: seasonality and trend, and applies different methods to make accurate predictions respectively, so that the model can more accurately capture the different factors affecting the power load, thereby improving the accuracy of the prediction;

[0026] 2. When using the hybrid expert decomposition and prediction model for prediction, the present invention adopts an analysis method that integrates local and global information for seasonal components, and introduces an innovative convolution mechanism within the extended layer and a residual block, thereby enhancing the ability of the hybrid expert decomposition and prediction model in predicting seasonal changes. For the prediction of the trend component, the damped trend method is used to improve its prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A schematic flow chart of the method is provided for the present invention. DETAILED DESCRIPTION

[0028] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0029] Example

[0030] like Figure 1 As shown, this embodiment provides a method for power load forecasting based on hybrid expert decomposition, comprising the following steps:

[0031] Step S1, obtaining and preprocessing the original data set, the original data set includes historical time series data of power load and weather data.

[0032] The preprocessing process specifically includes missing value processing and feature normalization. The missing values ​​and features are specifically the missing values ​​of the historical time series data of power load and the multi-dimensional weather information of the area collected on that day. Among them, the missing values ​​are filled by linear interpolation method; the Zscore scaling method is used to reduce the original data to a specific interval and perform normalization.

[0033] Step S2, performing correlation analysis on the pre-processed power load historical time series data and weather data, and selecting the corresponding power load historical time series data under the positive correlation dimension.

[0034] This step can obtain the influence of different factors on power load and present it as a correlation coefficient, including:

[0035] Since the historical time series data of power load all have corresponding weather dimension information after Zscore normalization, the random forest method is used to perform correlation analysis on the preprocessed historical time series data of power load and the corresponding weather dimension information, and the correlation coefficient between each weather dimension information and the corresponding historical time series data of power load is calculated;

[0036] Based on the correlation coefficient, the historical time series data of power load corresponding to the dimension that is positively correlated with the historical time series data of power load is selected, that is, the positive correlation dimension with a weight greater than 0 is selected.

[0037] Step S3, using the hybrid expert decomposition and prediction model to decompose the selected power load historical time series data into seasonal items and trend items, and make predictions respectively, and then splice different prediction results to obtain the final power load prediction result. The hybrid expert decomposition and prediction model includes a decomposition block, a seasonal item prediction block, a trend item prediction block, and a splicing block. Step S3 specifically includes:

[0038] Step S31, the decomposition block uses an expert hybrid strategy to decompose the selected power load historical time series data into seasonal items and trend items. Specifically, it includes:

[0039] (1) Decompose the historical time series data of power load from multiple dimensions;

[0040] (2) Decomposition stage: Receive the original time series data and input it into different "experts" in parallel. Each expert applies a specific decomposition method. For example, Fourier transform experts can capture periodic components, while wavelet transform experts are better at capturing sudden changes and non-periodic changes.

[0041] (3) Feature extraction stage: The output of each expert processing is a representation or component of the time series, which contains certain specific information or characteristics of the data. These outputs can be regarded as representations of the time series at different frequencies or scales;

[0042] (4) Ensemble learning phase: The outputs of these different experts are then integrated. This process involves weight learning, where the model learns the importance of different outputs to the final prediction results to determine how to most effectively combine this information. This step is implemented by a series of average filters of different sizes. The formalization is as follows:

[0043] X 趋势 =Softmax(L(x))*(F(x)) (1)

[0044] Among them, F(x) represents a set of average pooling filters, and Softmax(L(x)) represents the weighted fusion of the extracted trends.

[0045] According to formula (1), X 趋势 , decomposed into seasonal terms X through multi-scale s and trend term X t , and input them into the seasonal item prediction block and trend item prediction block respectively, and make predictions to get the final seasonal item prediction result Y s And the trend item prediction result Y t , and merged into the final result Y pred .

[0046] Step S32: the trend item prediction block predicts the trend item using the damped trend method to obtain a trend item prediction result.

[0047] In the trend term prediction substep, the damped trend method is used to obtain the trend period term. Specifically, the exponential smoothing method for trend prediction includes two recursions, namely S t and b t . Specifically expressed as:

[0048] S t =αX t +(1-α)(S t-1 +φb t-1 ) (2)

[0049] b t =β(S t -S t-1 )+(1-β)φb t-1 (3)

[0050] In the formula, S t represents the estimated value of the historical time series of power load, b t It represents the estimated value of the trend (slope) of the historical time series of power load. The subscripts t and t-1 represent the t time step and t-1 time step respectively. α is the level of the smoothing parameter (0≤α≤1), β is the smoothing parameter of the trend (0≤β≤1), and φ is a supplementary parameter.

[0051] If φ = 1, it is equivalent to the Holt trend method. The damped trend method (DTS) introduces a parameter, φ (0 < φ < 1) to "flatten" the trend at some point in the future for prediction. This parameter is in addition to the smoothing parameters α and β. φ flattens the trend of values ​​between 0 and 1 so that it approaches a constant at some point in the future. In fact, for any 0 < φ < 1, when h → ∞, the predicted value tends to S t +φb t / (1-φ)φ generally ranges from 0.8≤φ≤0.98.

[0052] Formula (2) is a horizontal equation, indicating that given S t-1 and b t-1 . S t It is a weighted average calculated from the predicted values ​​of the observed values ​​at time step t.

[0053] Formula (3) is a trend equation, indicating that b t Based on S t -S t-1 and b t-1 A weighted average of the previous trend estimates taken at time t. Here the function is no longer based only on the level, but on the trend. The forecast h steps away is equal to h times the previously estimated level plus the previously estimated trend. Therefore, the forecast is a linear function of h.

[0054] Step S33, the seasonal item prediction block uses multi-scale extended inner convolution to capture local features and global correlations and make predictions to obtain seasonal item prediction results.

[0055] The seasonal term prediction sub-step focuses on more complex seasonal components. After embedding the input sequence, multi-scale extended inner convolution is used to capture the local features and global correlation of the time series data. Branches of different scales process different basic patterns of the sequence. Subsequently, the results of different branches are merged to achieve full utilization of the sequence information. The seasonal term prediction sub-step is specifically expressed as:

[0056]

[0057] Y s,l = Function (Y s,l-1 ), l∈{1,2,…,N} (6)

[0058] Y s =Truncate( Projection (Y s,N )) (7)

[0059] Where, X zero ∈R O×drepresents a placeholder filled with zeros, Represents X s The embedding layer, Y s,l ∈R (I+O)×D represents the output of the lth multi-scale expansion layer within the convolutional layer, Y s ∈R O×d Represents the linear function projection Y s,N ∈R (I +O)×D Final forecast and truncation operations for the post-seasonal portion, represents the predicted value of the first layer, N represents the Nth embedding layer, I, O, and D represent the lth multi-scale convolutional layer, 0 value, and latitude, respectively.

[0060] The embedding method is replaced by a simpler complementary 0 strategy compared to the traditional encoder-decoder input. Using three-part embedded input, the embedding process is expressed as:

[0061]

[0062] Where TFE represents temporal feature encoding (e.g., one minute per hour, one minute per week, one day per week, one day per month, and one day per month), CVE represents static covariate information encoding, and VALUE represents value embedding.

[0063] The multi-scale convolutional layer consists of multiple branches, each using a different scale to capture the underlying temporal pattern of changes. In each branch, a local-global module is used to extract local features and global dependencies of the sequence. Specifically, the local module uses one-dimensional convolution to achieve downsampling after obtaining the corresponding single mode through average pooling. In addition, the global module aims to model the global correlation between the outputs of local modules. Although the self-attention mechanism is a common method to model global correlation, this method chooses a variant of dilated causal convolution - extended inner convolution as an alternative.

[0064] Dilated Inner Convolution As a variant of dilated causal convolution, equidistant convolution pads a sequence of length S with zero placeholders of length S-1 and sets its kernel to the same length. This means that a large convolution kernel can be used to measure the global correlation of the entire sequence. In addition, the convolution introduces jumps, resulting in gaps. When the dilation is set to 1, it functions as a regular causal convolution. When the dilation is set to 2, this design effectively expands the receptive field without significantly increasing the computational burden. Current generative prediction methods involve adding placeholders to the input sequence that do not contain actual sequence information in the second half of the sequence. Dilated inner convolution facilitates sequential reasoning of sequences by incorporating local feature information. To maintain the sequence length, the result of the dilated inner convolution is upsampled using a transposed convolution.

[0065] Step S34: the splicing block combines the trend item prediction result and the season item prediction result through a splicing operation to obtain the final power load prediction result.

[0066] Compared with other prediction methods, the power load prediction method provided in this embodiment has certain advantages, and the comparison results of statistical parameters are shown in Table 1-1 and Table 1-2. It can be seen from Table 1-1 and Table 1-2 that the method provided in this embodiment is compared with other combination models. Compared with several other current SOTA deep learning models, according to the comparison of MAE and MSE, it can also be clearly demonstrated that the method has higher prediction accuracy, which shows that the improved scheme is preferred.

[0067]

[0068] In the formula, m is the number of samples; y i is the true value; is the predicted value.

[0069] Table 1-1 Comparison results of statistical parameters of various prediction methods

[0070]

[0071]

[0072] Table 1-2 Comparison results of statistical parameters of various prediction methods (continued)

[0073]

[0074] Example 2

[0075] This embodiment provides a power load forecasting device based on hybrid expert decomposition, including a data acquisition and preprocessing module, a data screening module, and a decomposition and forecasting module.

[0076] The data acquisition and preprocessing module is used to acquire and preprocess the original data set, which includes the historical time series data of power load and weather data. This module can provide clean and standardized input data for power load forecasting. In some other embodiments, noise reduction processing can be performed by signal decomposition and combination methods.

[0077] The data screening module is used to perform correlation analysis on the preprocessed power load historical time series data and weather data, and select the corresponding power load historical time series data under the positive correlation dimension.

[0078] The decomposition and prediction module is used to decompose the selected power load historical time series data into seasonal items and trend items using a hybrid expert decomposition and prediction model, and predict them separately. After splicing different prediction results, the final power load prediction result is obtained. This decomposition enables the model to more accurately capture the different factors affecting the power load, thereby improving the accuracy of the prediction. Among them, the decomposition and prediction module includes a decomposition unit, a trend item prediction unit, a seasonal item prediction unit, and a splicing unit.

[0079] The decomposition unit is used to decompose the selected historical time series data of power load into seasonal items and trend items. It adopts an expert hybrid strategy, which mainly includes a decomposition stage, a feature extraction stage and an integrated learning stage. The integrated learning stage consists of a series of average filters of different sizes.

[0080] The trend item prediction unit is used to predict the trend item using the damped trend method to obtain the trend item prediction result. Through the fine modeling of the trend, the prediction model performs well in trend prediction, especially in the case of longer prediction steps.

[0081] The seasonal item prediction unit is used to use multi-scale extended inner convolution to capture local features and global correlations and make predictions to obtain seasonal item prediction results. On the one hand, the multi-scale extended inner convolution introduces the extended layer convolution mechanism and residual block when used to further enhance the model's ability to capture seasonal characteristics in power load data. This mechanism can improve the receptive field, produce holes, and improve prediction accuracy with lower complexity. On the other hand, by combining both local and global information, the entire prediction method and device can understand the data more comprehensively, thereby showing higher performance in predicting seasonal changes.

[0082] The splicing unit is used to splice the trend item prediction results and the seasonal item prediction results to obtain the final power load forecast result. This result takes into account the long-term dependence, periodic changes and distribution characteristics of the data, aiming to provide high-precision load forecasting.

[0083] The specific execution steps of each module and unit are the same as the execution steps of the method in Example 1, and will not be repeated here.

[0084] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for power load forecasting based on hybrid expert decomposition, characterized in that: The following steps are involved: S1, obtaining and preprocessing an original data set, wherein the original data set includes historical time series data of power load and weather data; S2, performing correlation analysis on the pre-processed historical time series data of power load and weather data, and selecting the corresponding historical time series data of power load under the positive correlation dimension; S3, using the hybrid expert decomposition and prediction model, the selected power load historical time series data is decomposed into seasonal items and trend items, and predictions are made separately. After splicing different prediction results, the final power load prediction result is obtained.

2. The power load forecasting method based on hybrid expert decomposition according to claim 1 is characterized in that: The hybrid expert decomposition and prediction model includes a decomposition block, a seasonal item prediction block, a trend item prediction block and a splicing block, and S3 specifically includes: The decomposition block decomposes the selected historical time series data of power load into seasonal items and trend items using an expert hybrid strategy; The trend item prediction block predicts the trend item using a damped trend method to obtain a trend item prediction result; The seasonal item prediction block uses multi-scale extended inner convolution to capture local features and global correlations and make predictions to obtain seasonal item prediction results; The splicing block splices the trend item prediction result and the season item prediction result to obtain a final power load prediction result.

3. The power load forecasting method based on hybrid expert decomposition according to claim 2 is characterized in that: The process of predicting the trend term using the damped trend method is expressed as: S t =αX t +(1-α)(S t-1 +φb t-1 ) b t =β(S t -S t-1 )+(1-β)φb t-1 In the formula, S t represents the estimated value of the historical time series of power load, b t represents the slope estimate of the historical time series of power load, the subscripts t and t-1 represent the t time step and t-1 time step respectively, α is the level of the smoothing parameter, β is the smoothing parameter of the trend, φ is the supplementary parameter, and X t For trend items.

4. The power load forecasting method based on hybrid expert decomposition according to claim 2 is characterized in that: The process of using multi-scale extended inner convolution to capture local features and global correlation and make predictions is expressed as: Y s，l =Function(Y s，l-1 ),l∈{1,2,...,N} Y s =Truncate(Projection(Y s，N )), Where, X zero ∈R O×d represents a placeholder filled with zeros, Represents X s The embedding layer, Y s,l ∈R (I +O)×D represents the output of the lth multi-scale expansion layer within the convolutional layer, Y s ∈R O×d Represents the linear function projection Y s,N ∈R (I+O)×D Final forecast and truncation operations for the post-seasonal portion, represents the predicted value of the first layer, N represents the number of embedding layers, I, O, and D represent the lth multi-scale convolutional layer, 0 value, and latitude, respectively.

5. The power load forecasting method based on hybrid expert decomposition according to claim 4 is characterized in that: The embedding process is expressed as: Where TFE represents temporal feature encoding, CVE represents static covariate information encoding, and VALUE represents value embedding.

6. The power load forecasting method based on hybrid expert decomposition according to claim 1 is characterized in that: The preprocessing process specifically includes missing value processing and feature normalization.

7. The power load forecasting method based on hybrid expert decomposition according to claim 6 is characterized in that: The S2 specifically includes: Based on the preprocessed historical time series data of power load, the corresponding weather dimension information is obtained; Perform correlation analysis on the preprocessed historical time series data of power load and the corresponding weather dimension information, and calculate the correlation coefficient between each weather dimension information and the corresponding historical time series data of power load; Based on the correlation coefficient, the electric power load historical time series data corresponding to the dimension that is positively correlated with the electric power load historical time series data is selected.

8. The power load forecasting method based on hybrid expert decomposition according to claim 7 is characterized in that: The correlation analysis was performed based on the random forest method.

9. A power load forecasting device based on hybrid expert decomposition, characterized in that: include: A data acquisition and preprocessing module, used to acquire and preprocess the original data set, wherein the original data set includes historical time series data of power load and weather data; A data screening module is used to perform correlation analysis on the pre-processed historical time series data of power load and weather data, and select the corresponding historical time series data of power load under the positive correlation dimension; as well as The decomposition and prediction module is used to decompose the selected historical time series data of power load into seasonal items and trend items using a hybrid expert decomposition and prediction model, and to make predictions for each of them. The final power load prediction result is obtained by splicing different prediction results.

10. The power load forecasting device based on hybrid expert decomposition according to claim 9, characterized in that: The decomposition and prediction module includes: A decomposition unit, used for decomposing the selected historical time series data of power load into seasonal items and trend items by using an expert hybrid strategy; A trend item prediction unit, used to predict the trend item by using a damped trend method to obtain a trend item prediction result; Seasonal item prediction unit, which uses multi-scale extended inner convolution to capture local features and global correlations and make predictions to obtain seasonal item prediction results; The splicing unit is used to splice the trend item prediction result and the season item prediction result to obtain a final power load prediction result.

Images