A fine-grained power data prediction method, system, device and medium
By using pre-trained networks of EMD, multi-branch BiLSTM and DLSTM layers in power data prediction, the problem of data leakage and low accuracy in fine-grained power data prediction is solved, and higher prediction accuracy and real-time decision support is achieved.
Patent Information
- Application Number
- CN202411194550.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-29
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2044-08-29
AI Technical Summary
The prior art faces the problems of data leakage and low prediction accuracy in fine-grained prediction of power loads, especially in multi-dimensional time series prediction, and fails to fully utilize the information of relevant variables.
Using pre-trained networks, including empirical modal decomposition (EMD) decomposition layer, multi-branch BiLSTM layer and deep LSTM (DLSTM) layer, preliminary prediction is carried out through EMD stationary processing and multi-branch BiLSTM, and preliminary prediction results are carried out through DLSTM layer. Correction to avoid data breaches and improve prediction accuracy.
It effectively avoids data leakage, improves the accuracy of fine-grained power data prediction, can better capture power trends, and provide real-time and detailed decision-making basis for power production and allocation.
Smart Images

Figure CN118709866B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of industrial power energy consumption prediction, and in particular to a fine-grained power data prediction method, system, equipment and medium. Background Art
[0002] China's huge population base determines the huge demand for electricity. The generation, transportation, and use of electricity are carried out simultaneously, and the electricity demand and population distribution in different places are not the same. All of these have brought great difficulties to meeting people's electricity needs. How to accurately predict power data from all levels is the key to solving the problem of power energy allocation. In particular, the fine-grained prediction of power load can avoid unplanned production and unreasonable allocation of electricity by providing control and decision-making support for power allocation and management, which can greatly alleviate the pressure on regional electricity consumption.
[0003] Fine-grained prediction of power load faces huge challenges. In the research of coarse-grained prediction of power load data, there are many achievements. The amount of data for coarse-grained prediction of the same data will be much less than that for fine-grained prediction, and the information and details of the data will also be less than those of fine-grained prediction. Coarse-grained prediction has lower requirements on the accuracy and details of prediction results than fine-grained prediction. The coarse-grained test set cannot judge the quality of prediction results from a more microscopic and detailed level, so coarse-grained prediction can often easily achieve good results. However, the actual power consumption situation changes rapidly, and coarse-grained prediction cannot meet the needs of power load prediction well due to the loss of data details and larger granularity. To provide real-time and detailed decision-making basis for power production and allocation, fine-grained prediction with smaller granularity is required. Compared with coarse-grained prediction, fine-grained prediction has finer data granularity, which means that the data has more detailed information and more abnormal points. And for power data, power trends are more difficult to capture in fine-grained conditions. All these have brought great difficulties to fine-grained prediction. At present, there are some common problems in the academic community regarding time series prediction:
[0004] For example, according to a large number of literature surveys, some papers have the problem of test data leakage when conducting time series prediction; and when using EMD for multi-dimensional time series prediction, many of them only use the dimension of data labels, treating multi-dimensional data as single variable data for prediction, and fail to make full use of the information of other related variables. Summary of the invention
[0005] The purpose of the present invention is to provide a fine-grained power data prediction method, system, device and medium, which can avoid data leakage and improve the accuracy of data prediction.
[0006] To achieve the above object, the present invention provides the following solutions:
[0007] A fine-grained power data prediction method, comprising:
[0008] Get raw power data;
[0009] The original power data is input into the pre-trained network for optimization training, and the trained network is determined as a prediction model; the prediction model is used to perform fine-grained power data prediction to obtain a final prediction result; the pre-trained network includes an EMD decomposition layer, a multi-branch BiLSTM layer and a DLSTM layer connected in sequence; wherein the multi-branch BiLSTM layer includes a forward LSTM layer, a backward LSTM layer, an input layer and an output layer, and the forward LSTM layer and the backward LSTM layer are respectively connected to the input layer and the output layer;
[0010] The specific process of inputting the original power data into the pre-training network for optimization training is as follows:
[0011] Inputting the original power data into the EMD decomposition layer, using the empirical mode decomposition method to perform a stabilization process on the power time series, and obtaining a plurality of intrinsic mode function components and a residual component;
[0012] Send each component to the multi-branch BiLSTM layer for prediction to obtain a preliminary prediction result;
[0013] The preliminary prediction results corresponding to each component are feature fused with the original power data, and the fused feature data are input into the DLSTM layer for re-prediction, so as to correct the preliminary prediction results.
[0014] Optionally, before inputting the original power data into the pre-training network for optimization training, it also includes: preprocessing the original power data to obtain preprocessed data; the preprocessing method includes resampling and missing value filling.
[0015] Optionally, the specific process of the preprocessing is: resampling the original power data at a set granularity, and filling the missing values of the sampled data according to the mean filling method.
[0016] Optionally, before each component is respectively sent to the multi-branch BiLSTM layer for prediction, it also includes: using a sliding window to divide the components into a supervised learning data format.
[0017] The present invention also provides a fine-grained power data prediction system, comprising:
[0018] Data acquisition module, used to obtain raw power data;
[0019] A fine-grained power data prediction module is used to input the original power data into a pre-trained network for optimization training, and determine the trained network as a prediction model; the prediction model is used to perform fine-grained power data prediction to obtain a final prediction result; the pre-trained network includes an EMD decomposition layer, a multi-branch BiLSTM layer and a DLSTM layer connected in sequence; wherein the multi-branch BiLSTM layer includes a forward LSTM layer, a backward LSTM layer, an input layer and an output layer, and the forward LSTM layer and the backward LSTM layer are respectively connected to the input layer and the output layer;
[0020] The specific process of inputting the original power data into the pre-training network for optimization training is as follows:
[0021] Inputting the original power data into the EMD decomposition layer, using the empirical mode decomposition method to perform a stabilization process on the power time series, and obtaining a plurality of intrinsic mode function components and a residual component;
[0022] Send each component to the multi-branch BiLSTM layer for prediction to obtain a preliminary prediction result;
[0023] The preliminary prediction results corresponding to each component are feature fused with the original power data, and the fused feature data are input into the DLSTM layer for re-prediction, so as to correct the preliminary prediction results.
[0024] The present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the above-mentioned fine-grained power data prediction method.
[0025] The present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program implements the fine-grained power data prediction method as described above when executed by a processor.
[0026] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:
[0027] The present invention discloses a fine-grained power data prediction method, system, device and medium, the method comprising obtaining original power data; inputting the original power data into a pre-trained network for optimization training, and determining the trained network as a prediction model; the prediction model is used to predict fine-grained power data to obtain a final prediction result; the pre-trained network comprises an EMD decomposition layer, a multi-branch BiLSTM layer and a DLSTM layer connected in sequence; wherein the EMD decomposition layer is used to decompose an irregular wave into a plurality of IMFs and a residual form; the multi-branch BiLSTM layer comprises a forward and a backward two-layer LSTM, and the forward layer and the backward layer are respectively connected to the input layer and the output layer for preliminary prediction; the DLSTM layer comprises a plurality of LSTM layers, and a fully connected layer connected to each LSTM layer, for correcting the preliminary prediction result. The present invention can improve the accuracy of data prediction and avoid leakage of test data when training the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0029] Figure 1 This is a flow chart of fine-grained power data prediction of the present invention;
[0030] Figure 2 This is a structural diagram of LSTM and BiLSTM in this embodiment;
[0031] Figure 3 is a structural diagram of the DLSTM in this embodiment;
[0032] Figure 4 This is a prediction result diagram in this embodiment;
[0033] Figure 5 This is the distribution diagram of the first 360 data after resampling in this embodiment;
[0034] Figure 6 This is the EMD decomposition result diagram in this embodiment. DETAILED DESCRIPTION
[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0036] The purpose of the present invention is to provide a fine-grained power data prediction method, system, device and medium, which can avoid data leakage and improve the accuracy of data prediction.
[0037] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0038] The present invention provides a fine-grained power data prediction method, comprising:
[0039] Step 100: Obtain raw power data.
[0040] Step 200: Input the original power data into Figure 1 The pre-trained network shown in the figure is optimized and trained, and the trained network is determined as a prediction model; the prediction model is used to perform fine-grained power data prediction to obtain the final prediction result; the pre-trained network includes an EMD decomposition layer, a multi-branch BiLSTM layer and a DLSTM layer connected in sequence; wherein the EMD decomposition layer is used to decompose an irregular wave into a number of IMFs and a residual form; the multi-branch BiLSTM layer includes two LSTM layers, a forward layer and a backward layer, and the forward layer and the backward layer are respectively connected to the input layer and the output layer for preliminary prediction; the DLSTM layer includes multiple LSTM layers, and a fully connected layer connected to each LSTM layer, which is used to correct the preliminary prediction result.
[0041] Based on the above technical solution, the important components in each step are described in detail.
[0042] Explanation of professional terms:
[0043] LSTM: Long Short-Term Memory;
[0044] BiLSTM: Bidirectional Long Short-Term Memory;
[0045] DLSTM: Deep Long Short-Term Memory;
[0046] EMD: Empirical Mode Decomposition;
[0047] IMF: Intrinsic Mode Function;
[0048] RNN: Recurrent Neural Network;
[0049] GAP: Global active power;
[0050] preprocessing: preprocessing;
[0051] Pre-forecast: prediction;
[0052] correct: correction.
[0053] For the multi-branch BiLSTM network layer: Traditional methods predict multi-dimensional data by often splicing them into a large matrix and then inputting the whole into a neural network for prediction. However, data of different dimensions have different distribution patterns, and using the same neural network for prediction ignores the differences in the distribution of different data. Therefore, this solution invented a multi-branch BiLSTM, the number of branches can change dynamically with the data dimension to keep the same as the data dimension. When predicting, different data is input into the corresponding branch for prediction. In this way, each branch learns its unique distribution pattern based on the corresponding data, making the final prediction result more accurate.
[0054] For multi-feature fusion datasets: In order to make full use of the features of the dataset and prevent the model from overfitting, this solution uses the multi-branch BiLSTM preliminary prediction results as one of the new features and fuses them with the original dataset to form a multi-feature fusion dataset. This facilitates the next step of DLSTM for re-prediction.
[0055] For re-prediction using the DLSTM layer: After the data is predicted by the multi-branch BiLSTM, the prediction effect is poor in some places where the data fluctuates greatly. Therefore, this solution invents a DLSTM re-prediction method to re-predict the data after the multi-branch BiLSTM prediction, so as to achieve the effect of correcting some of the poorly predicted areas. Compared with different LSTMs, DLSTM has a deeper feature extraction level, can perform deeper feature extraction on multi-dimensional data, and better mine data information.
[0056] Among them, Figure 2 As shown in the figure, the left side is the overall structure of BiLSTM, and the right side is the internal structure of LSTM. , Represents the forward layer and the backward layer The latent vector at time t is and They represent the input and output of BiLSTM at time t respectively.
[0057] Therefore, in order to solve the problem of fine-grained prediction of electric energy in industry, the specific process is as follows:
[0058] Step 1: Data preprocessing method, follow the steps 1.1~1.2 below:
[0059] Step 1.1: Resample the input data: If the data granularity is too small, the data will be more easily affected by various factors and distributed more irregularly and randomly, making prediction more difficult. In actual power management and dispatch, predictions in hours can meet the needs, so this embodiment resamples the data in one-hour granularity.
[0060] Step 1.2: Fill missing values with mean value: After resampling, there are still missing values in the data, so it is necessary to fill the data to ensure that the subsequent prediction can proceed normally. There are many ways to deal with missing values, such as overall mean filling, median filling, linear interpolation, etc. Comprehensive experimental results show that the effect of using mean filling is the best.
[0061] Step 2: Time series stabilization method based on EMD decomposition:
[0062] In time series prediction, for non-stationary time series, they are generally converted into stationary time series before prediction. EMD is essentially a stabilization process for data series or signals, which is suitable for decomposing nonlinear and non-stationary time series. Therefore, this embodiment uses EMD as a means to stabilize power load data. The basic idea of EMD is to decompose an irregular wave into several IMFs and a residual. The basic process of decomposition is performed according to the following steps 2.1 to 2.4:
[0063] Step 2.1: Based on the original signal All the maximum and minimum points form the upper envelope of the original signal and lower envelope ;
[0064] Step 2.2: Find the mean of the upper and lower envelopes to form the mean envelope ,Right now:
[0065] ,
[0066] Step 2.3: Subtract the mean envelope from the original signal to get the intermediate signal ,Right now:
[0067] ,
[0068] Step 2.4: Determine the intermediate signal Check whether the two constraints of IMF are met. If so, the intermediate signal is an IMF component. If not, the signal is the original signal and step 1 is repeated. Step 4, until the intermediate signal obtained in step 3 satisfies the two constraints of IMF.
[0069] Step 3: Multi-branch BiLSTM preliminary prediction:
[0070] In the time series prediction model, LSTM has better long-term memory function than ordinary RNN. It effectively solves the problem of gradient disappearance, can handle time series problems well, and has achieved success in many fields. The LSTM model can capture long-distance dependencies, but it only uses past information as input and hidden layer, and ignores future information. BiLSTM is an extension of the unidirectional LSTM model. It is composed of two layers of LSTM, the forward layer and the backward layer, and the input layer and the output layer are connected respectively. The BiLSTM model is Figure 3 shown.
[0071] Suppose the i-th component obtained after the dataset label Y is decomposed is , each All components have their own distribution law. Merge them into a matrix and send them to BiLSTM for training. A single BiLSTM model is difficult to learn the rules of each component. Therefore, this embodiment uses a multi-branch BiLSTM to train each Each branch of BiLSTM trains a BiLSTM model separately. Perform independent training and learning, compared to a single BiLSTM for all Prediction, multi-branch BiLSTM can achieve more accurate prediction results.
[0072] The basic process of multi-branch BiLSTM prediction is performed according to the following steps 3.1~3.4:
[0073] Step 3.1: Data format conversion: The i-th component obtained after EMD decomposition It cannot be directly fed into BiLSTM for operation. It is necessary to use a sliding window to split the components into supervised learning data formats that BiLSTM can train. Assume that the size of the sliding window is w+1, that is, using The first w historical data predict the current value. yes The jth value of The kth training sample :
[0074] ,
[0075] In the formula, { } is the characteristic of the data, is the prediction target. predictions, Figure 3 In that is , that is The predicted value of .
[0076] Step 3.2: BiLSTM prediction: The sample set is framed by sliding window It will be sent to the BiLSTM of the corresponding branch for prediction. and The combination can better capture the dependencies between time series nodes, thereby improving the effect of time series prediction. The calculation process is as follows:
[0077] ,
[0078] ,
[0079] ,
[0080] in, and Respectively and right The latent vector obtained by calculation is for and The result after splicing is and predicted results . Apply multi-branch BiLSTM to all The prediction results are concatenated to obtain a new component sequence :
[0081] ,
[0082] In the formula, is the forecast result of the ith IMF component, Represents the prediction result of the residual component.
[0083] Step 3.3: Combine the prediction components to get the preliminary prediction result: According to the characteristics of EMD, all the components of the original sequence decomposed by EMD are added together, and the resulting sequence is equal to the original sequence. Therefore, the sum of all the components in S is equal to the shape of Y. By adding up each component of :
[0084] ,
[0085] ,
[0086] Among them, R is the residual component in the EMD decomposition result.
[0087] Step 4: Error correction based on DLSTM prediction results: The current model only uses the information of the Y dimension in the dataset D, and the performance of the model is too dependent on Y. Label Y is easily affected by other random factors and has outliers or large fluctuations. At this time, the model will learn the influence of these noises or unreasonable abnormal patterns, which is easy to cause overfitting. Preliminary prediction results Often, the actual value will be very different in places with large fluctuations and abnormal points. If the model input has other features besides Y, the model will fully consider each feature when calculating, and the overall judgment will not be affected by the abnormal value of a certain feature. Therefore, training the model with a multi-feature data set can effectively alleviate the problem of model overfitting.
[0088] In order to make full use of the characteristics of the dataset , to prevent the model from overfitting. This embodiment will initially predict the results Fuse with other features in the original data set to form a multi-feature fusion data set . And send it to DLSTM for prediction again as a correction to the initial prediction result. The specific steps are as follows 4.1~4.3:
[0089] Step 4.1: Multiple features are integrated to form a new data set: In this example, the initial prediction results As one of the new features, it is fused with the original dataset to form a multi-feature fusion dataset :
[0090] ,
[0091] Step 4.2: DLSTM prediction:
[0092] BiLSTM predicts single-dimensional data differently. It is a multidimensional data set, which contains more information and requires a deeper network for feature extraction. Therefore, a multi-layer DLSTM is used here. Compared with the single-layer LSTM, DLSTM can mine multidimensional data sets. Deeper information can achieve better prediction results. The structure of DLSTM is as follows Figure 3 shown.
[0093] in, Similarly, the data needs to be segmented into supervised learning format after sliding window segmentation before it can be sent to DLSTM for prediction. Suppose the kth sample after segmentation is , let the sample The corresponding predicted target load value is , For the kth sample input to DLSTM:
[0094] ,
[0095] ,
[0096] ,
[0097] in, Indicates time t Then the sample set Send it to DLSTM for prediction as the preliminary prediction result An error correction, the output result The final prediction result.
[0098] The definitions or concepts involved in the above detailed steps are explained below:
[0099] Fine-grained prediction: Empirical mode decomposition (EMD), a new method for processing non-stationary signals. The time-frequency analysis method based on EMD is suitable for the analysis of nonlinear and non-stationary signals as well as linear and stationary signals, and the analysis of linear and stationary signals also reflects the physical meaning of the signal better than other time-frequency analysis methods.
[0100] Resampling: refers to the process of converting a time series from one frequency to another. It is a method of reprocessing the original sample and a convenient method for resampling and frequency conversion of regular time series data.
[0101] Furthermore, a specific embodiment is provided.
[0102] In this embodiment, the data set comes from the power load data of the UCI database (individual household electric power consumption). The sampling frequency of the data set is every minute, the time span of data collection is 47 months, and the number is 2,075,259 in total, of which missing values are recorded as '?', accounting for about 1.25%. Each data contains a total of 8 dimensions including the time dimension, among which global active power (GAP) is the object of prediction.
[0103] The prediction effect of the EMD-BiLSTM model obtained in this embodiment is as follows: Figure 4 As shown, the final prediction result after correction can be seen Significantly better than initial predictions It is closer to the true value, proving that the DLSTM multi-feature prediction module of this embodiment plays a great role.
[0104] Considering that if the data granularity is too small, the data will be more easily affected by various factors and distributed more irregularly and randomly, making prediction more difficult. In actual power management and allocation, forecasts in hours can meet the needs. Therefore, this embodiment resamples the data originally in one-minute granularity with one hour as the granularity, and the amount of data after resampling becomes 34,589. There will still be missing values after resampling. There are many ways to deal with missing values, such as overall mean filling, median filling, linear interpolation, etc. Comprehensive experimental results show that the effect of using mean filling is the best. The GAP distribution of the first 360 data after resampling is as follows Figure 5 As shown in the figure, there are no missing values after resampling and mean filling. However, the overall distribution pattern is still not obvious and fluctuates greatly.
[0105] In order to better test the performance of the model, this embodiment divides the data set into a training set and a test set, accounting for 80% and 20% of the total, respectively. Considering that the numerical values of each dimension of the data vary greatly, it is not conducive to the convergence of model training. Therefore, the data is normalized to normalize the original data values to between [0,1] as follows:
[0106] ,
[0107] In the formula, x is a dimension other than time in the dataset D, min(x) and max(x) are the minimum and maximum values of this dimension respectively. is the normalized value.
[0108] Since the data input to the model has been normalized, the power load forecast data output by the model is also normalized. In order to compare the forecast results with the original data on the same dimension, the output data needs to be denormalized as follows:
[0109] ,
[0110] The data label Y is decomposed into 8 IMF components and one residual component by EMD, such as Figure 6 As shown. Figure 5 Compared with the original data, the component series after EMD decomposition is more stable and the rules are more obvious.
[0111] In the model of this embodiment, the number of hidden units unit1 and unit2 in the first layer of the BiLSTM module and DLSTM are extremely important hyperparameters, which have a great impact on the performance of the model. When unit1 and unit2 are 128 and 256 respectively, the overall model performs best. Therefore, the number of hidden layer units in the first layer of BiLSTM and DLSTM is set to 128 and 256 respectively. The DLSTM model structure used in the experiment is 3 LSTM layers plus a fully connected layer. The activation function used by the 3-layer LSTM is tanh, and the number of hidden units is 256, 128, and 64 respectively. The activation function used in the fully connected layer is linear.
[0112] EMD-BiLSTM-DLSTM will produce preliminary prediction results in the middle After that, the preliminary prediction results will be processed and predicted again to get the final prediction results. When the hyperparameters are set to the best, the model outputs the preliminary prediction results of the active power GAP, the final prediction results and the actual GAP. Figure 4 As shown. Obviously It is closer to the true value, proving that the DLSTM multi-feature prediction module of this embodiment plays a great role.
[0113] As can be seen from the above, the present invention first uses empirical mode decomposition (EMD) to decompose the power time series to make the data tend to be stable, and then uses a bidirectional long short-term memory network (BiLSTM) to perform preliminary prediction of the time series. Finally, the preliminary prediction results are fused with other features in the original data set and input into a deep long short-term memory network (DLSTM) to train the model, and the output results are used as corrections to the preliminary prediction results. This method is used for fine-grained prediction of electricity and can provide control and decision-making support for the production, allocation and management of electricity. This method can be used in scenarios such as power station energy allocation and factory power cost assessment.
[0114] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.
[0115] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the core idea of the present invention. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.
Claims
1. A fine-grained power data prediction method, characterized in that: include: Get raw power data; The original power data is input into the pre-trained network for optimization training, and the trained network is determined as a prediction model; the prediction model is used to perform fine-grained power data prediction to obtain a final prediction result; the pre-trained network includes an EMD decomposition layer, a multi-branch BiLSTM layer and a DLSTM layer connected in sequence; wherein the multi-branch BiLSTM layer includes a forward LSTM layer, a backward LSTM layer, an input layer and an output layer, and the forward LSTM layer and the backward LSTM layer are respectively connected to the input layer and the output layer; the number of branches of the multi-branch BiLSTM layer is the same as the dimension of the data, and each output component of the EMD decomposition layer is individually connected to a BiLSTM model branch; the DLSTM layer includes a plurality of LSTM layers connected in sequence and a fully connected layer; The specific process of inputting the original power data into the pre-training network for optimization training is as follows: The original power data is input into the EMD decomposition layer, and the power time series is stabilized by using the empirical mode decomposition method to obtain a plurality of intrinsic mode function components and a residual component; Send each component to the multi-branch BiLSTM layer for prediction to obtain a preliminary prediction result; The preliminary prediction results corresponding to each component are subjected to feature fusion with the original power data, and the fused feature data are input into the DLSTM layer for re-prediction, so as to correct the preliminary prediction results; Before each component is sent to the multi-branch BiLSTM layer for prediction, the method further includes: using a sliding window to divide the components into a supervised learning data format.
2. The fine-grained power data prediction method according to claim 1, characterized in that: Before the original power data is input into the pre-training network for optimization training, the method also includes: pre-processing the original power data to obtain pre-processed data; the pre-processing method includes resampling and missing value filling.
3. The fine-grained power data prediction method according to claim 2, characterized in that: The specific process of the preprocessing is: resampling the original power data at a set granularity, and filling the missing values of the sampled data according to the mean filling method.
4. A fine-grained power data prediction system, characterized in that: include: Data acquisition module, used to obtain raw power data; A fine-grained power data prediction module is used to input the original power data into a pre-trained network for optimization training, and determine the trained network as a prediction model; the prediction model is used to perform fine-grained power data prediction to obtain a final prediction result; the pre-trained network includes an EMD decomposition layer, a multi-branch BiLSTM layer and a DLSTM layer connected in sequence; wherein the multi-branch BiLSTM layer includes a forward LSTM layer, a backward LSTM layer, an input layer and an output layer, and the forward LSTM layer and the backward LSTM layer are respectively connected to the input layer and the output layer; the number of branches of the multi-branch BiLSTM layer is the same as the dimension of the data, and each output component of the EMD decomposition layer is individually connected to a BiLSTM model branch; the DLSTM layer includes a plurality of LSTM layers connected in sequence and a fully connected layer; The specific process of inputting the original power data into the pre-training network for optimization training is as follows: The original power data is input into the EMD decomposition layer, and the power time series is stabilized by using the empirical mode decomposition method to obtain a plurality of intrinsic mode function components and a residual component; Send each component to the multi-branch BiLSTM layer for prediction to obtain a preliminary prediction result; The preliminary prediction results corresponding to each component are subjected to feature fusion with the original power data, and the fused feature data are input into the DLSTM layer for re-prediction, so as to correct the preliminary prediction results; Before each component is sent to the multi-branch BiLSTM layer for prediction, the method further includes: using a sliding window to divide the components into a supervised learning data format.
5. An electronic device, characterized in that: It comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the fine-grained power data prediction method according to any one of claims 1-3.
6. A computer-readable storage medium, characterized in that: It stores a computer program, which, when executed by a processor, implements the fine-grained power data prediction method according to any one of claims 1 to 3.
Citation Information
Patent Citations
Epidemic situation prediction analysis method and system for epidemic disease intelligent monitoring system
CN114596966A
Wind power output prediction method and system
CN114861513A
Short-term power load prediction method
CN117709393A
