Load prediction method for electricity consumption information acquisition terminal
Through the combination of the dual-channel input structure and the autoencoder, the problem of insufficient adaptability of the existing power load prediction method in different scenarios is solved, and a higher precision load prediction is achieved.
Patent Information
- Application Number
- CN202410085105.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-20
- Publication Date
- 2025-07-29
AI Technical Summary
The existing power load prediction methods require a lot of manual experience construction, and cannot adapt to different power scenarios, and fail to effectively utilize future data characteristics, resulting in insufficient prediction accuracy.
The dual-channel input structure is used to process historical data and predicted time feature data respectively, the convolutional neural network and gated loop unit are used to extract features, and the prediction results are generated in combination with the autoencoder to adapt to the feature dimensions and time points changes in different scenarios.
It improves the accuracy and adaptability of load prediction, and can achieve accurate prediction in different electric usage scenarios.
Smart Images
Figure CN120387010A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of terminals, and in particular, to a load forecasting method for an electricity consumption information collection terminal. Background Art
[0003] Currently, power load curve forecasting technology has been widely applied, which mainly includes empirical forecasting methods, statistical forecasting methods, and artificial intelligence-based forecasting methods. Empirical forecasting methods mainly infer future load conditions based on empirical formulas, statistical formulas, or expert rules, including time series analysis methods, regression analysis methods, etc. Such methods usually require historical load data and various related data for a long time. After multiple statistical regressions, key indicators are filtered out to establish a forecasting model. Statistical forecasting methods use probability and statistical methods to infer future load conditions and can be applied to various forecasting types such as long-term forecasting, short-term forecasting, and medium-term forecasting. For example, appropriate statistical methods such as ARIMA models, regression models, neural networks, simulated annealing, etc. can be selected to find the optimal forecasting method, and then modeling analysis is performed based on historical data.
[0004] The above two methods require a large amount of artificial experience to construct a forecasting algorithm, and the constructed forecasting system often cannot be applied to different power scenarios. The artificial intelligence-based forecasting method is a new forecasting method that has emerged in recent years. It uses artificial intelligence technologies such as neural networks, support vector machines, and deep learning to analyze historical load data and establish a forecasting model. This method can effectively process complex non-linear data and uncertain factors and has good self-adaptability, robustness, and generalization ability. There are an endless stream of forecasting models constructed based on deep learning methods, but most studies only use historical data for training when constructing models and do not consider the influence of future factors on load data. A small number of studies consider this point and combine historical data and future data into a data matrix to obtain a forecasting model, but do not consider that the proportion of historical data in a data sample is much higher than that of future data. This way of training the model makes the information of future data not fully learned. At the same time, most models do not consider that the number of features that can be obtained in different electricity consumption scenarios is different, and the number of forecasting time points is fixed, so they cannot be flexibly applied to various scenarios. Summary of the Invention
[0005] In view of the deficiencies and defects existing in the prior art, the present invention proposes a load prediction method for an electricity consumption information collection terminal. By means of a dual-channel input structure, this method can better learn the characteristics of historical data and future data, provide more reasonable characteristic information for the prediction part, and contribute to the improvement of prediction accuracy. At the same time, the input part of the model is designed to be able to adjust for characteristic information of different dimensions without affecting the internal processing logic of the model, and can be adapted to different electricity consumption scenarios to achieve load prediction.
[0006] The object of the present invention can be achieved by the following technical solutions:
[0007] A load prediction method for an electricity consumption information collection terminal, the method comprising the following steps:
[0008] Step 1: Data collection: Collect load data and relevant information affecting load changes according to the electricity consumption scenario.
[0009] Step 2: Data processing: First, clean the collected data and perform standardization processing; sort the data set in ascending order of time, use the first 70% of the data as the training set, and the last 30% of the data as the test set; then construct training samples for the data set respectively, and the training samples include a historical data matrix, a data feature matrix at the prediction moment, and the actual load data at the prediction moment.
[0010] Step 3: Model construction: The model is divided into three parts: feature extraction, feature fusion, and data generation. The feature extraction part adopts a dual-channel structure design. The two channels use different design methods for learning historical data and data features at the prediction moment, and use the splicing fusion method to fuse the feature information extracted by the two channels. Finally, an autoencoder is used to generate the load data at the prediction moment.
[0011] Step 4: Model training: Based on the training data set in Step 2, use the Adam optimization function to repeatedly train the model, and use the training set to verify the loss value between the predicted value and the actual value until the weight value of the load prediction model is determined after reaching the expected range.
[0012] Step 5: Electricity consumption prediction: Obtain the actual data in the electricity consumption scenario that is the same as the data features in the training samples in Step 2, and construct the same composition structure; then input it into the load prediction model obtained in Step 4, and the result of the output layer is the predicted load result at the next time point.
[0013] Furthermore, the training samples in step 2 are constructed based on the characteristics of the model. Each row of the historical data matrix represents all the data features at a time point, and they are arranged in ascending order of time points from top to bottom. Each column of the data matrix at the prediction time represents all the data features at a time point, and they are arranged in ascending order of time points from left to right. The actual load data at the prediction time is arranged from top to bottom in ascending order of time. Among them, the data at the prediction time is continuous with the historical data in terms of time, and the actual load data at the prediction time is the same as the data at the prediction time in terms of time.
[0014] Furthermore, in the model feature extraction part of step 3, the historical data processing channel first uses a convolutional neural network (CNN) to extract the relevant information of the feature data at a time point. The number of convolutional kernels is M, and the size is (1*N). This method will not be affected by the information of other time points, so that the original time information is retained between the features extracted by each convolutional kernel. Then, the batch normalization layer is used to standardize the feature data to accelerate the training process of the model. Finally, the gated recurrent unit (GRU) is used to learn the temporal feature information between the feature sequences.
[0015] Furthermore, in the model feature extraction part of step 3, the prediction time data feature processing channel uses an autoencoder with neurons decreasing layer by layer to compress the feature information at a time point, and finally flattens the compressed feature data of all time points into a feature sequence.
[0016] Furthermore, in the data generation part of step 3, an autoencoder is constructed using a combination method of neurons first decreasing and then increasing, which can further fuse the feature information through the encoding and decoding operations of the fused feature sequence, and generate the expected data features based on the historical data features and prediction time feature data extracted from the two channels. Finally, the sigmod function is used to generate the load prediction data.
[0017] Furthermore, the input structure of the historical data processing channel proposed in step 3 has the same convolutional kernel size as the number of historical data features. Based on the characteristics of the data samples constructed in step 2, it can be adapted only by modifying the convolutional kernel size according to the change in the number of feature quantities in the prediction scenario, without affecting the internal data processing logic of the model. At the same time, the number of prediction time points can be freely selected, and only the number of neurons in the output layer needs to be modified.
[0018] Furthermore, the method is an end-to-end prediction model constructed based on deep learning. Only by inputting the specified data into the model can the prediction result of the future load be obtained.
[0019] Advantageous technical effects of the present invention: A new load prediction model is proposed based on deep learning technology. Based on the characteristics of a large amount of historical data and scarce feature data at the prediction moment, a dual-channel input structure is constructed to process historical data and feature data at the prediction moment respectively. While deepening the learning depth of historical data, more feature information at the prediction moment can be retained, providing a more reasonable proportion of feature information for the prediction part and increasing the accuracy of the model's load prediction. The feature data of the two channels are spliced and fused using the idea of feature fusion, and the load data at future moments are accurately predicted with the help of an autoencoder. To adapt to the changes in the input feature dimensions and the number of prediction time points in different scenarios, based on the adjustable characteristics of the neural network neuron parameters, a design for the model to adapt to different-dimensional input and output is added, enabling it to be applicable to load prediction in multiple scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 It is a schematic diagram of the deep learning model for load prediction of the present invention.
[0021] Figure 2 It is a flowchart of the model training of the present invention.
[0022] Figure 3 It is a flowchart of the model load prediction of the present invention DETAILED DESCRIPTION OF THE INVENTION
[0023] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not limit the present invention.
[0024] The present invention will be further described below with reference to the accompanying drawings.
[0025] As Figure 3 shown, a method for predicting the load of an electricity consumption terminal based on deep learning is as follows:
[0026] Step 1: Select relevant features that affect load changes. The available feature information varies in different electricity consumption scenarios. In addition to the key historical load data, data factors such as temperature, humidity, air pressure, wind speed, light intensity, season, time, electricity bill, type of electricity consumers, and social policies all directly or indirectly affect the change of electricity load. It is necessary to reasonably select data features with strong influencing factors according to the characteristics of the electricity consumption scenario and add special influencing factor features when necessary.
[0027] Step 2: Collect load data. According to the time interval between the predicted load data, periodically collect the relevant features selected in Step 1 to form the original feature data arranged in ascending order of time.
[0028] Step 3: Data sorting and constructing a dataset. In the process of collecting the original dataset, there are cases of missing or incorrect data, so cleaning operations need to be performed on it. Based on the original dataset, according to the time sequence, the first 70% of the data is used as the training set, and the last 30% of the data is used as the test set. Each sample data in the two datasets is divided into a historical data matrix H, a data matrix P at the prediction moment, and the actual load data R at the prediction moment. Among them, each row of the historical data matrix H represents all the feature data (x1, x2, x3…x n , s) collected at a certain moment, including the load data s, and is arranged in ascending order of time from top to bottom. Each column of the data matrix P at the prediction moment contains all the feature data (y1, y2, y3…y m ) except the load data s, and is arranged in ascending order of time from left to right. The actual load data R is collected at the same continuous moment as the prediction data matrix P and forms a one-dimensional matrix (s1, s2, s3…s m ) in chronological order.
[0029] Step 4: Constructing a load prediction model. The prediction model proposed in this patent is as Figure 1 shown. It proposes a dual-channel structure to process historical data and data at the prediction moment respectively. Using the idea of feature fusion, the feature information extracted from the two channels is spliced and fused. Then, an autoencoder is used to generate prediction feature information. Finally, the sigmod function is used to obtain the load prediction result. Among them, for the channel processing historical data, first, a convolutional network with a convolutional kernel size of (1*N) is used to learn the correlation information between all the feature data in a moment. Then, the extracted information is optimized through a batch normalization layer (BN) to increase the training speed of the model. The convolutional kernel size is the same as the number of features of the input historical data. This method can better retain the temporal correlation between feature sequences. Then, a gated recurrent unit (GRU) is used to learn the temporal features between the feature data. This channel can perform in-depth learning on the complex historical data to better obtain feature data with more complete representation information. For the prediction data processing channel, an autoencoder with a compression function is used for learning. Since the data at the prediction time point is scarce, this method can better retain the feature information. After compression, a new feature sequence is formed by flattening. The processing methods of the two channels enable a more balanced feature sequence to be obtained in the feature fusion part. In the subsequent autoencoder network, prediction data features can be generated based on the rich feature information. The number of neurons on both sides of the autoencoder structure is relatively large, and the number of neurons in the middle is relatively small. This structure can be regarded as an encoding and decoding operation. First, the feature information is compressed, and then decoded into the desired feature data, which conforms to the idea of generating predicted load data. Finally, the sigmod function is used to calculate the predicted load data.
[0030] Step 5: Training the load prediction model. AsFigure 2 As shown, according to steps 1, 2, 3, and 4, the data set and prediction model are constructed. First, the weight parameters of the model are initialized, and then the training set data is input into the model for training. The Adam optimization function is used to optimize the model parameters to gradually reduce the loss value. The loss value is calculated using the L1 regularization loss function, and the formula is as follows:
[0031]
[0032] where represents the load data output by the prediction model, y represents the actual load data value in the data training set sample, and λ and w are hyperparameter constants that control the size of the regularization term.
[0033] Then, the loss value of the model prediction result is verified using the test set data. If the loss value does not meet the requirements, the model is retrained using the training data set to optimize the network weights until the loss value obtained using the test set reaches the expected target. Finally, the network weight parameters of the prediction model are obtained.
[0034] Step 6: Input actual data to obtain the prediction result. Collect actual data according to the composition structure of the data sample in step 3, and input the data into the prediction model obtained in step 5 to obtain the predicted load data at future time points.
[0035] The above embodiments are illustrative of the specific implementation manners of the present invention, rather than limitations on the present invention. Those skilled in the relevant technical fields can make various transformations and changes without departing from the spirit and scope of the present invention to obtain corresponding equivalent technical solutions. Therefore, all equivalent technical solutions should be included in the patent protection scope of the present invention.
Claims
1. A load forecasting method for an electricity consumption information acquisition terminal, characterized in that, It includes the following steps: Step 1: Data collection: Collect load data and relevant information affecting load changes according to the electricity consumption scenario; Step 2: Data processing: First, clean the collected data and perform standardization processing; Sort the data set in ascending order of time, take the first 70% of the data as the training set, and the last 30% of the data as the test set; Then construct training samples for the data set respectively. The training samples include a historical data matrix, a data feature matrix at the prediction moment, and the actual load data at the prediction moment; Step 3: Model construction: The model is divided into three parts: feature extraction, feature fusion, and data generation. The feature extraction part adopts a dual-channel structure design. The two channels use different design methods for learning historical data and data features at the prediction moment, and use the splicing fusion method to fuse the feature information extracted by the two channels. Finally, an autoencoder is used to generate the load data at the prediction moment; Step 4: Model training: Based on the training data set in Step 2, use the Adam optimization function to repeatedly train the model, and use the training set to verify the loss value between the predicted value and the actual value until the expected range is reached, and then determine the weight value of the load prediction model; Step 5: Electricity consumption prediction: Obtain the actual data in the electricity consumption scenario that is the same as the data features in the training samples in Step 2, and construct the same composition structure; Then input it into the load prediction model obtained in Step 4, and the result of the output layer is the predicted load result for the next time point.
2. The load prediction method for an electricity consumption information acquisition terminal according to claim 1, characterized in that, It is described that: The training samples in Step 2 are constructed based on the characteristics of the model. Each row of the historical data matrix represents all the data features of a time point, and is arranged in ascending order of time from top to bottom; Each column of the data matrix at the prediction moment represents all the data features of a time point, and is arranged in ascending order of time from left to right; The actual load data at the prediction moment is arranged from top to bottom in ascending order of time; Among them, the data at the prediction moment is continuous with the historical data in time, and the actual load data at the prediction moment is the same as the data at the prediction moment in time.
3. A load forecasting method for an electricity consumption information collection terminal according to claim 1, characterized in that, It is described that: In the model feature extraction part of Step 3, the historical data processing channel first uses a convolutional neural network (CNN) to extract the relevant information of the feature data of a time point. The number of convolutional kernels is M, and the size is (1*N). This method will not be affected by the information of other time points, so that the features extracted by each convolutional kernel retain the original time information. Then use the batch normalization layer to standardize the feature data to accelerate the training process of the model. Finally, use the gated recurrent unit (GRU) to learn the time feature information between the feature sequences.
4. A load forecasting method for an electricity consumption information acquisition terminal according to claim 1, characterized in that It is described that: In the model feature extraction part of Step 3, the prediction moment data feature processing channel uses an autoencoder with gradually decreasing neurons to compress the feature information of a time point, and finally flattens the compressed feature data of all time points into a feature sequence.
5. A load forecasting method for an electricity consumption information acquisition terminal according to claim 1, characterized in that, It is described that in the data generation part of step 3, an autoencoder is constructed using a combination of neurons that first decrease and then increase, which can further fuse feature information for the encoding and decoding operations of the fused feature sequence, generate expected data features based on the historical data features extracted from the dual channels and the feature data at the prediction moment; finally, the sigmod function is used to generate the load prediction data.
6. A load forecasting method for an electricity consumption information collection terminal according to claim 1, characterized in that, It is described that the input structure of the historical data processing channel proposed in step 3 has the same convolutional kernel size as the number of historical data features. Based on the characteristics of the data samples constructed in step 2, it can be adapted by simply modifying the convolutional kernel size according to the change in the number of features in the prediction scenario without affecting the internal data processing logic of the model; at the same time, the number of prediction time points can be freely selected by simply modifying the number of neurons in the output layer.
7. A load forecasting method for an electricity consumption information collection terminal according to claim 1, characterized in that, The method described is an end-to-end prediction model constructed based on deep learning, and the prediction result of the future load can be obtained by simply inputting the specified data into the model.