Power supply load prediction method based on convolutional neural network
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUILIN UNIV OF ELECTRONIC TECH
- Filing Date
- 2022-06-30
- Publication Date
- 2026-06-26
Smart Images
Figure CN115358437B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of smart grid load forecasting and analysis technology, specifically a power supply load forecasting method based on convolutional neural networks. Background Technology
[0002] Currently, the increasing maturity of power information technology has created broad prospects for the application of power big data. Power data is diverse in type and broad in scope, spanning the entire power production and consumption process. Power demand forecasting is of great significance for power construction planning, grid dispatch and control, and power market assessment.
[0003] As a crucial component of energy management systems, power system load forecasting's accuracy directly impacts subsequent power grid safety verification analyses, significantly affecting power grid dynamic state estimation, load dispatching, and reducing generation costs. Due to the inherent uncertainty and complexity of loads, accurate load characteristic analysis and forecasting model development are key to improving forecasting precision.
[0004] Existing prediction models mainly focus on shallow learning, which has limited ability to approximate complex functions with limited samples and computing units. It is difficult to extract deep features of load sequences, which limits the generalization performance of the models and hinders further improvement in prediction accuracy. Furthermore, traditional electricity demand forecasting is based on simple electricity information and lacks external data correlation analysis, resulting in a lack of prediction of changing trends beyond the fitted patterns and low prediction accuracy. Summary of the Invention
[0005] (a) Technical problems to be solved
[0006] To address the shortcomings of existing technologies, this invention provides a power supply load prediction method based on convolutional neural networks. It constructs a convolutional neural network power prediction algorithm, which improves the efficiency of massive data processing in the power prediction process, comprehensively considers related information such as temperature, and overcomes the problem of excessive reliance on personal experience in the prediction process. As a result, it can reduce personnel requirements, simplify the operation process, and make it more convenient to use.
[0007] (II) Technical Solution
[0008] To achieve the above objectives, the present invention provides a power supply load prediction method based on convolutional neural networks, comprising the following steps:
[0009] S1. Obtain historical load data of the power system in the preset area, and preprocess abnormal data in the historical load data;
[0010] S2. Analyze and quantify the factors affecting power load, and standardize the corrected load data. Specifically, use the normalization formula to normalize the load data to [0, 1] so that the load data are at the same order of magnitude, thereby accelerating the convergence speed of the neural network.
[0011] S3. Construct a multi-scale information fusion convolutional neural network model and divide the obtained dataset into two parts: a training set and a validation set. The training set is used to train the model and the validation set is used to validate the training results.
[0012] S4. Select a training set vector sequence of fixed length as the model input, and use the actual load at a fixed time after the input vector as the training target of the model output. Train the model until convergence through multiple iterations, use the validation set to verify the trained model, and adjust the model parameters by comparing the accuracy and error of the test set and the validation set. After multiple training sessions, select the model with the best performance on the validation set as the training result.
[0013] S5. Run the model in a real-world environment. When there is a large deviation between the predicted and actual values, add the latest data to the training set and retrain the model to obtain the latest and most practical convolutional neural network model for power prediction.
[0014] S6. Using a convolutional neural network model, extract the proximity, periodicity, and trend features of electricity consumption based on the complex nonlinear mapping relationship between hourly, daily, and weekly electricity consumption and external data such as temperature and holidays, and construct a convolutional neural network for electricity consumption prediction.
[0015] S7. Preprocess the input layer data on electricity consumption, temperature, and holidays. After normalizing the data into dimensionless relative quantities, input the massive data into the electricity prediction convolutional neural network in step S6 and initialize the weights and biases of the multi-channel convolutional neural network.
[0016] S8. The input data is processed layer by layer through a convolutional neural network. The weights and biases of each layer are adjusted based on the backpropagation algorithm of the error gradient. Training is stopped after the set number of iterations is reached. The prediction result is obtained by inputting the test sample set.
[0017] Preferably, in step S1, the data sample of the historical power load data of a certain region comes from the data acquisition and monitoring control system of that region, and the abnormal data preprocessing is carried out by using analytical analysis and correction methods to identify bad data and then supplement the missing data.
[0018] Preferably, the factors affecting the power load in step S2 include temperature, weather characteristics, and date type, and these factors are quantified according to their degree of influence on the load.
[0019] Preferably, the process of constructing the multi-scale information fusion convolutional neural network model in step S3 is as follows:
[0020] 1) Introducing causal logic constraints enhances the representation of time series features. Causal convolution operations only perform forward convolution and do not acquire information from future moments.
[0021] 2) Utilize multi-scale convolution to describe the interrelationships between time-domain data of different lengths;
[0022] 3) Utilize residual network structures to improve network depth and prediction accuracy.
[0023] Preferably, in step S7, the above three types of data are transformed into normalized dimensionless data, and the calculation formula is as follows:
[0024]
[0025] In the formula:
[0026] y: Normalized data value;
[0027] y max =1;
[0028] y min =-1;
[0029] x max : The maximum value of the data before normalization;
[0030] x min Minimum value of data before normalization.
[0031] Preferably, in step S8, the weights are initialized using a Xavier normal distribution:
[0032]
[0033]
[0034] E(w) = 0, ensuring that the weights follow a uniform distribution with a mean of 0;
[0035] Where E represents the mean, Var represents the variance, and n j n represents the number of nodes in the j-th layer. j+1 This represents the number of nodes in the (j+1)th layer.
[0036] Preferably, the formula for calculating the convolutional layer in step S8 is:
[0037]
[0038] n = 1, 2, ..., C0
[0039] In the above formula, y mLet x be the output of the m-th convolutional layer, a be the activation function of the convolutional layer, and x be the output of the m-th convolutional layer. j For the j-th channel input, w m For the m-th convolutional kernel, b m C0 is the bias value, and C0 is the total number of convolution kernels.
[0040] Preferably, in step 2), the relationship between temporal data of different lengths is described by multi-scale convolution as follows: the convolution kernel size, scale factor and network depth are determined by the length of the input sequence, so that the product of the three is greater than the length of the input sequence.
[0041] Preferably, in step 3), the residual network structure is used to improve the network depth and prediction accuracy as follows: a residual structure is adopted, which is composed of multiple residual blocks stacked together, wherein one residual block includes multi-scale information fusion convolution, weight normalization, activation function, and Dropout structure.
[0042] (III) Beneficial Effects
[0043] This invention provides a power supply load prediction method based on convolutional neural networks. It has the following beneficial effects:
[0044] 1. This invention provides a power load prediction method based on convolutional neural networks. By modeling power load data and related influencing factors as a time series problem, a multi-scale information fusion convolutional neural network is proposed to predict short-term power load. Causal logic constraints are introduced to enhance the expression of time series features, and multi-scale convolution is used to describe the relationship between time-domain data of different lengths. Finally, a residual network structure is designed to improve network depth and prediction accuracy. When faced with a very long one-dimensional input sequence, the method proposed in this invention can still extract multi-dimensional features of the entire sequence, which can effectively improve the accuracy of power load prediction.
[0045] 2. This invention provides a power supply load prediction method based on convolutional neural networks. This method targets the complex nonlinear mapping relationship between hourly, daily, and weekly power consumption and external data such as temperature and holidays. It extracts the proximity, periodicity, and trend features of power consumption, constructs a power consumption prediction convolutional neural network, and then converts the three types of data into dimensionless relative quantities through normalization. Massive data is input into the convolutional neural network for layer-by-layer calculation. The backpropagation algorithm of error gradient is used to adjust the weights and biases in each layer of the network. After reaching the required number of training iterations, training stops and the power consumption prediction result is output.
[0046] 3. This invention provides a power supply load prediction method based on convolutional neural networks. This method constructs a convolutional neural network power supply prediction algorithm according to three types of data: power consumption, temperature, and holidays. It improves the efficiency of massive data processing in the power supply prediction process, comprehensively considers related information such as temperature, and overcomes the problem of excessive reliance on personal experience in the prediction process. As a result, it can reduce personnel requirements, simplify the operation process, and make it more convenient to use. Attached Figure Description
[0047] Figure 1 This is a flowchart of the power supply load prediction method based on convolutional neural networks according to the present invention. Detailed Implementation
[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0049] Example:
[0050] like Figure 1 As shown, this embodiment of the invention provides a power supply load prediction method based on a convolutional neural network, including the following steps:
[0051] S1. Obtain historical load data of the power system in a certain region, and preprocess abnormal data in the historical load data;
[0052] S2. Analyze and quantify the factors affecting power load, and standardize the corrected load data. Specifically, use the normalization formula to normalize the load data to [0, 1] so that the load data are at the same order of magnitude, thereby accelerating the convergence speed of the neural network.
[0053] S3. Construct a multi-scale information fusion convolutional neural network model and divide the obtained dataset into two parts: a training set and a validation set. The training set is used to train the model and the validation set is used to validate the training results.
[0054] S4. Select a training set vector sequence of fixed length as the model input, and use the actual load at a fixed time after the input vector as the training target of the model output. Train the model until convergence through multiple iterations, use the validation set to verify the trained model, and adjust the model parameters by comparing the accuracy and error of the test set and the validation set. After multiple training sessions, select the model with the best performance on the validation set as the training result.
[0055] S5. Run the model in a real-world environment. When there is a large deviation between the predicted and actual values, add the latest data to the training set and retrain the model to obtain the latest and most practical convolutional neural network model for power prediction.
[0056] S6. Using a convolutional neural network model, extract the proximity, periodicity, and trend features of electricity consumption based on the complex nonlinear mapping relationship between hourly, daily, and weekly electricity consumption and external data such as temperature and holidays, and construct a convolutional neural network for electricity consumption prediction.
[0057] S7. Preprocess the input layer data on electricity consumption, temperature, and holidays. After normalizing the data into dimensionless relative quantities, input the massive data into the electricity prediction convolutional neural network in step S6 and initialize the weights and biases of the multi-channel convolutional neural network.
[0058] S8. The input data is processed layer by layer through a convolutional neural network. The weights and biases of each layer are adjusted based on the backpropagation algorithm of the error gradient. Training is stopped after the set number of iterations is reached. The prediction result is obtained by inputting the test sample set.
[0059] The process of constructing the multi-scale information fusion convolutional neural network model in step S3 is as follows:
[0060] 1) Introducing causal logic constraints enhances the representation of time series features. Causal convolution operations only perform forward convolution and do not acquire information from future moments.
[0061] 2) Utilize multi-scale convolution to describe the interrelationships between time-domain data of different lengths;
[0062] 3) Utilize residual network structures to improve network depth and prediction accuracy.
[0063] In step 2), multi-scale convolution is used to describe the relationship between temporal data of different lengths: the kernel size, scale factor and network depth are determined by the length of the input sequence, so that the product of the three is greater than the length of the input sequence.
[0064] Step 3) utilizes a residual network structure to improve network depth and prediction accuracy: a residual structure is adopted, which is composed of multiple residual blocks stacked together. One of the residual blocks includes multi-scale information fusion convolution, weight normalization, activation function, and Dropout structure.
[0065] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A power supply load prediction method based on convolutional neural networks, characterized in that: Includes the following steps: S1. Obtain historical load data of the power system in the preset area, and preprocess abnormal data in the historical load data; S2. Analyze and quantify the factors affecting power load, and standardize the corrected load data. Specifically, use the normalization formula to normalize the load data to [0, 1] so that the load data are at the same order of magnitude, thereby accelerating the convergence speed of the neural network. S3. Construct a multi-scale information fusion convolutional neural network model and divide the obtained dataset into two parts: a training set and a validation set. The training set is used to train the model and the validation set is used to validate the training results. S4. Select a training set vector sequence of fixed length as the model input, and use the actual load at a fixed time after the input vector as the training target of the model output. Train the model until convergence through multiple iterations, use the validation set to verify the trained model, and adjust the model parameters by comparing the accuracy and error of the test set and the validation set. After multiple training sessions, select the model with the best performance on the validation set as the training result. S5. Run the model in a real-world environment. When there is a large deviation between the predicted and actual values, add the latest data to the training set and retrain the model to obtain the latest and most practical convolutional neural network model for power prediction. S6. Using a convolutional neural network model, extract the proximity, periodicity, and trend features of electricity consumption based on the complex nonlinear mapping relationship between hourly, daily, and weekly electricity consumption and external data such as temperature and holidays, and construct a convolutional neural network for electricity consumption prediction. S7. Preprocess the input layer data on electricity consumption, temperature, and holidays. After normalizing the data into dimensionless relative quantities, input the massive data into the electricity prediction convolutional neural network in step S6 and initialize the weights and biases of the multi-channel convolutional neural network. S8. The input data is processed layer by layer through a convolutional neural network. The weights and biases of each layer are adjusted based on the backpropagation algorithm of the error gradient. Training is stopped after the set number of iterations is reached. The prediction result is obtained by inputting the test sample set. In step S1, the data sample of the historical power load data of a certain region comes from the data acquisition and monitoring control system of that region, and the abnormal data preprocessing is to identify bad data by analytical analysis and correction methods, and then supplement the missing data. The factors affecting power load in step S2 include temperature, weather characteristics, and date type, and these factors are quantified according to their degree of influence on the load. The process of constructing the multi-scale information fusion convolutional neural network model in step S3 is as follows: 1) Introducing causal logic constraints enhances the representation of time series features. Causal convolution operations only perform forward convolution and do not acquire information from future time points. 2) Utilize multi-scale convolution to describe the interrelationships between time-domain data of different lengths; 3) Utilize residual network structures to improve network depth and prediction accuracy.
2. The power supply load prediction method based on convolutional neural networks according to claim 1, characterized in that: In step S7, the above three types of data are transformed into normalized dimensionless data, and the calculation formula is as follows: ； In the formula: y: Normalized data value; y max ＝1； y min ＝-1； x max : The maximum value of the data before normalization; x min : Minimum value of data before normalization; In step S8, the weights are initialized using a Xavier normal distribution. ； ； E(w) = 0, ensuring that the weights follow a uniform distribution with a mean of 0; Where E represents the mean, Var represents the variance, and n j Let n represent the number of nodes at level j. j+1 This represents the number of nodes in the (j+1)th layer; The formula for calculating the convolutional layer in step S8 is: ； n = 1, 2, ..., C0; In the above formula, y m Let x be the output of the m-th convolutional layer, a be the activation function of the convolutional layer, and x be the output of the m-th convolutional layer. j For the j-th channel input, w m For the m-th convolutional kernel, b m C0 is the bias value, and C0 is the total number of convolution kernels; In step 2), multi-scale convolution is used to describe the relationship between temporal data of different lengths: the kernel size, scale factor and network depth are determined by the length of the input sequence, so that the product of the three is greater than the length of the input sequence. Step 3) utilizes a residual network structure to improve network depth and prediction accuracy as follows: a residual structure is adopted, which is composed of multiple stacked residual blocks, one of which includes multi-scale information fusion convolution, weight normalization, activation function, and Dropout structure.