Power load prediction method based on deep learning

Abstract

The invention relates to the field of power load prediction, in particular to a power load prediction method based on deep learning. Comprising the steps of collecting and preprocessing historical power load data to obtain power load data, further constructing a time sequence energy entropy network, calculating an energy value, calculating an energy entropy sequence matrix according to the energy value, and extracting energy entropy features from the energy entropy sequence matrix; fusing the energy entropy features to generate a comprehensive feature vector; dynamically adjusting a weight matrix and a bias term of the time sequence energy entropy network, and predicting a power load; and evaluating a prediction result, optimizing the time sequence energy entropy network, performing power load prediction by using the optimized time sequence energy entropy network, and generating an analysis report of the prediction result. The problems that a current deep learning model is slow in response to complex changes of power load data, and results are lack of accuracy when fluctuations on different time scales are processed are solved; and when the power load data fluctuates abnormally or changes violently, the prediction precision is greatly reduced.

Description

Technical Field

[0001] The present invention relates to the field of power load forecasting, and in particular to a power load forecasting method based on deep learning. Background Art

[0002] In the field of power load forecasting, traditional methods mostly use statistical models based on time series analysis and basic machine learning techniques, such as ARIMA, support vector machine (SVM) and simple neural network models. The method can handle certain nonlinear and time-dependent power load data, but with the increase in the complexity and diversity of power load data, the analysis ability of a single time scale has gradually become insufficient. In recent years, deep learning methods have been gradually applied to power load forecasting, especially models such as long short-term memory networks (LSTM) and convolutional neural networks (CNN), which have significantly improved the accuracy and stability of prediction by extracting complex patterns and nonlinear features in power load data. However, most of the methods rely on single-scale data processing, and it is difficult to effectively capture the dynamic change characteristics of power load data at different time scales.

[0003] Although the application of deep learning in power load forecasting has made significant progress, there are still some technical problems that need to be solved: the current deep learning model can only process data features on a single time scale, which limits the ability to accurately capture the long-term and short-term dependencies in power load data. Since power load data essentially has multi-scale characteristics, ignoring the above characteristics will cause the prediction model to respond slowly to the complex changes in power load data, especially when dealing with fluctuations on different time scales, the performance of the prediction model is often not ideal. In addition, many existing prediction models lack the ability to adjust parameters in real time when dealing with dynamic changes and sudden fluctuations in power load data, resulting in a significant decrease in prediction accuracy when power load data fluctuates abnormally or changes drastically. The above technical problems limit the widespread application of existing deep learning models in actual power load forecasting, and a more flexible and accurate prediction model is urgently needed to solve them. Summary of the invention

[0004] The present invention provides a power load forecasting method based on deep learning, which solves the problem that the current deep learning model can only process data features on a single time scale, which limits the ability to accurately capture the long-term and short-term dependencies of the power load data, resulting in the slow response of the forecasting model to the complex changes of the power load data, especially when dealing with fluctuations on different time scales, the results of the forecasting model lack accuracy; the existing forecasting model lacks the ability to adjust parameters in real time when dealing with dynamic changes and sudden fluctuations in power load data, resulting in a technical problem that the forecasting accuracy is greatly reduced when the power load data undergoes abnormal fluctuations or drastic changes.

[0005] A power load forecasting method based on deep learning of the present invention specifically includes the following technical solutions:

[0006] A power load forecasting method based on deep learning, comprising the following steps:

[0007] S1: Collect and preprocess historical power load data to obtain power load data; construct a time series energy entropy network based on the power load data, calculate the energy value, and calculate the energy entropy based on the energy value; based on the energy entropy, form an energy entropy sequence matrix, and extract energy entropy features from the energy entropy sequence matrix;

[0008] S2: Fusion of energy entropy features to generate a comprehensive feature vector; introduction of an energy entropy dynamic adjustment mechanism to dynamically adjust the weight matrix and bias term of the time series energy entropy network; prediction of the power load based on the comprehensive feature vector, the dynamically adjusted weight matrix, and the dynamically adjusted bias value; evaluation of the prediction results, optimization of the time series energy entropy network, use of the optimized time series energy entropy network to predict the power load, and generation of an analysis report of the prediction results.

[0009] Preferably, the S1 specifically includes:

[0010] The time series energy entropy network consists of an input layer, a feature extraction layer, a feature fusion layer and an output layer.

[0011] Preferably, the S1 specifically includes:

[0012] The time series energy entropy network receives power load data through the input layer and calculates the energy value. The formula is as follows:

[0013]

[0014] Among them, E t is the energy value of the power load value at time point t; N is the number of power load values ​​at time point t; x t，i is the ith power load value at time point t; x t，j is the jth power load value at time point t; i is the power load value index; j is the index variable of the power load value summation operation; t is the time index, indicating the current time point; α is the attenuation coefficient; max(x t ) is the maximum value of all power load values ​​at time point t; is an exponential decay factor.

[0015] Preferably, the S1 specifically includes:

[0016] The time scale and time window length are introduced. At each time scale, the energy entropy is arranged in chronological order to obtain an energy entropy sequence. Starting from the starting point of the energy entropy sequence, the time window is intercepted according to the time window length, and the time window slides backward, one time step at a time, until the entire energy entropy sequence is covered, and the energy entropy sequence matrix is ​​obtained. The energy entropy in each time window represents a row of the energy entropy sequence matrix.

[0017] Preferably, the S1 specifically includes:

[0018] The time series energy entropy network performs convolution operation through the feature extraction layer to extract energy entropy features from the energy entropy sequence matrix.

[0019] Preferably, the S2 specifically includes:

[0020] The time series energy entropy network fuses the energy entropy features through the feature fusion layer to form a comprehensive feature vector.

[0021] Preferably, the S2 specifically includes:

[0022] The output layer of the time series energy entropy network introduces an energy entropy dynamic adjustment mechanism. The energy entropy dynamic adjustment mechanism adjusts the weight matrix and bias value of the time series energy entropy network in real time according to the change of energy entropy to obtain a dynamically adjusted weight matrix and a dynamically adjusted bias value.

[0023] Preferably, the S2 specifically includes:

[0024] The power load is predicted using the comprehensive eigenvector, the dynamically adjusted weight matrix and the dynamically adjusted bias value; the prediction formula is as follows:

[0025]

[0026] in, is the predicted value of the time series energy entropy network at time point t+1; is the dynamically adjusted weight matrix; is the dynamically adjusted bias value; ⊙ is the element-by-element multiplication operation; F t is the comprehensive feature vector at time point t; ||·|| is the Euclidean length of the calculation vector.

[0027] Preferably, the S2 specifically includes:

[0028] The predicted value is denormalized and compared with the actual power load data. The prediction result is evaluated by calculating the prediction error. When the prediction error is greater than the preset error threshold, the time series energy entropy network is optimized and adjusted. The power load is predicted using the optimized time series energy entropy network until the prediction error is no greater than the error threshold, and an analysis report of the prediction result is generated.

[0029] The beneficial effects of the technical solution of the present invention are:

[0030] 1. By introducing the time series energy entropy network (TEENet), high-precision prediction of power load data is achieved. TEENet uses feature extraction, fusion and dynamic adjustment mechanisms to capture the nonlinear characteristics and dynamic changes in power load data, improve the accuracy and robustness of the time series energy entropy network, and provide more accurate power load forecasting results in complex power systems through in-depth mining and multi-dimensional analysis of energy entropy features, providing reliable data support for power dispatching and demand management.

[0031] 2. The weight matrix and bias value of the time series energy entropy network are adjusted in real time through the energy entropy dynamic adjustment mechanism (EDAM), so that the time series energy entropy network can adapt to the dynamic changes of power load data, have strong anti-interference ability, make the prediction results stable, effectively deal with the uncertainty factors in the power system, and improve the applicability and reliability of the time series energy entropy network.

[0032] 3. During the evaluation of the prediction results, the prediction error is calculated and compared with the set error threshold. When the prediction error exceeds the error threshold, the time series energy entropy network is retrained or parameter optimization is performed through the back propagation algorithm and gradient descent method to continuously improve the prediction accuracy of the time series energy entropy network, providing a strong guarantee for the stable operation and decision-making optimization of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flow chart of a method for power load forecasting based on deep learning described in the present invention. DETAILED DESCRIPTION

[0034] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0035] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0036] The following is a detailed description of a method for power load forecasting based on deep learning provided by the present invention in conjunction with the accompanying drawings.

[0037] Refer to the attached Figure 1 , which shows a flow chart of a method for power load forecasting based on deep learning provided by an embodiment of the present invention, the method comprising the following steps:

[0038] S1: Collect and preprocess historical power load data to obtain power load data; construct a time series energy entropy network based on the power load data, calculate the energy value, and calculate the energy entropy based on the energy value; based on the energy entropy, form an energy entropy sequence matrix, and extract energy entropy features from the energy entropy sequence matrix.

[0039] First, historical power load data is collected and preprocessed to obtain power load data. The preprocessing process includes data cleaning, normalization, and missing value filling. Specifically, data cleaning uses a z-score-based method, normalization uses a Min-Max scaling method, and missing value filling uses a time series-based linear interpolation method. The preprocessing process uses existing technologies and will not be described in detail here.

[0040] The power load data is the power load value at a series of time points, which is a collection of power load values. Based on the power load data, a time series energy entropy network (TEENet) is constructed. Specifically, the input layer of TEENet receives the power load data and converts the power load data into an energy entropy sequence matrix; the feature extraction layer extracts energy entropy features from the energy entropy sequence matrix through convolution operations; the feature fusion layer fuses the energy entropy features and outputs a comprehensive feature vector; the output layer combines the comprehensive feature vector, the dynamically adjusted weight matrix, and the dynamically adjusted bias value to generate a prediction result to realize power load prediction. TEENet can adapt to changes in power load data and has anti-interference capabilities and the ability to quickly adjust the response speed.

[0041] First, the input layer of TEENet receives power load data and calculates the energy value of the power load value at a point in time. The energy value is expressed in scalar form as an intensity indicator of the power load value at a point in time, quantifying the time point characteristics of the power load data and indicating the distribution and change pattern of the power load data. This provides a basis for the subsequent calculation of energy entropy.

[0042] The energy value calculation formula is as follows:

[0043]

[0044] Among them, E t is the energy value of the power load value at time point t; N is the number of power load values ​​at time point t; x t，i is the ith power load value at time point t; x t，j is the jth power load value at time point t; i is the power load value index; j is the index variable of the power load value summation operation; t is the time index, indicating the current time point; α is the attenuation coefficient, which is the adjustment parameter of the exponential attenuation factor and is set according to the specific implementation scenario through expert experience method; max(x t ) is the maximum value of all power load values ​​at time point t; is an exponential decay factor.

[0045] After obtaining the energy value of the power load value at the time point, the energy entropy of the power load value at the time point is calculated. The energy entropy is expressed in scalar form and is used to quantify the time point characteristics of the power load data. The larger the energy entropy, the more complex the energy distribution of the power load data at the time point. The energy entropy calculation formula enhances the sensitivity of the time series energy entropy network to the changes in the energy distribution of the power load data by introducing logarithmic functions, trigonometric functions and normalization processing, and can better capture the nonlinear characteristics in the power load data.

[0046] The energy entropy calculation formula is as follows:

[0047]

[0048] Among them, H t is the energy entropy of the power load value at time point t; E t is the energy value of the power load value at time point t; T is the number of time points; t′ is the index variable of the summation operation of the time points;

[0049] In order to capture the dynamic change characteristics of power load data at different time scales, the time scale k is introduced to represent the time range traced back from the current time point. By setting the time window length, TEENet analyzes the change of energy entropy at multiple time scales. The time scale and time window length should be adjusted according to the actual application scenario. At each time scale k, the energy entropy is arranged in chronological order to form an energy entropy sequence. Starting from the starting point of the energy entropy sequence, the time window is intercepted according to the set time window length, and the time window slides backward, one time step at a time, until the entire energy entropy sequence is covered, and the energy entropy sequence matrix is ​​obtained. The energy entropy in each time window represents a row of the energy entropy sequence matrix, which represents the energy entropy sequence in the corresponding time window at a specific time scale.

[0050] The feature extraction layer of TEENet extracts energy entropy features from the energy entropy sequence matrix through convolution operations. The energy entropy feature extraction formula is as follows:

[0051]

[0052] Among them, f t，k It is the energy entropy feature extracted at time point t on time scale k, expressed in the form of a vector; ReLU is the activation function; It is the convolution kernel weight matrix, which is used to perform convolution operation on the energy entropy sequence and extract the energy entropy features; is the energy entropy sequence at time point t on time scale k; is the convolution bias vector at time point t on time scale k; and In the process of constructing the time series energy entropy network, random initialization (Gaussian distribution or uniform distribution) is used to generate the initial value, and in the process of training the time series energy entropy network, the back propagation algorithm and the gradient descent method are used to update it. Random initialization, back propagation algorithm, and gradient descent method are all existing technologies and will not be described in detail here; is the energy entropy sequence matrix Performs exponential operation element by element; ⊙ is element by element multiplication operation; ||·|| is used to calculate the Euclidean length of a vector.

[0053] S2: Fusion of energy entropy features to generate a comprehensive feature vector; introduction of an energy entropy dynamic adjustment mechanism to dynamically adjust the weight matrix and bias term of the time series energy entropy network; prediction of the power load based on the comprehensive feature vector, the dynamically adjusted weight matrix, and the dynamically adjusted bias value; evaluation of the prediction results, optimization of the time series energy entropy network, use of the optimized time series energy entropy network to predict the power load, and generation of an analysis report of the prediction results.

[0054] In TEENet, the feature fusion layer fuses the energy entropy features extracted by the feature extraction layer to form a comprehensive feature vector for the final power load forecast. The fusion process combines element-by-element multiplication and logarithmic operations to retain the detailed information of the time scale and control the impact of the time scale on the forecast result through the weight matrix.

[0055] The fusion formula is as follows:

[0056]

[0057] Among them, F t is the comprehensive feature vector at time point t; W kIt is the weight matrix of the kth time scale. In the process of constructing the time series energy entropy network, random initialization (Gaussian distribution or uniform distribution) is used to generate the initial value. In the process of training the time series energy entropy network, it is updated by the back propagation algorithm and the gradient descent method. ∈ is a smoothing parameter, which is used to reduce the uncertainty in logarithmic operations and is set according to expert experience based on the specific implementation scenario. K is the number of time scales.

[0058] The output layer of TEENet introduces the Energy Entropy Dynamic Adjustment Mechanism (EDAM) to cope with the dynamic changes of power load data. The EDAM mechanism adjusts the weight matrix and bias value of the time series energy entropy network in real time according to the change in energy entropy, so that the time series energy entropy network can adapt to sudden changes in power load data. The adjusted weight matrix and bias value are adjusted nonlinearly and dynamically through the exponential function to prevent numerical instability.

[0059] The weight matrix adjustment formula is as follows:

[0060]

[0061] in, is a dynamically adjusted weight matrix, which is the EDAM mechanism for W k The weight matrix after dynamic adjustment; △H t is the change in energy entropy at time point t. Assuming that the change in energy entropy at the initial moment is zero, the subsequent energy entropy change is calculated as △H t =H t -H t-1 ;H t-1 is the energy entropy of the power load value at time point t-1; λ is the weight threshold parameter, which is used to prevent numerical overflow and is set according to expert experience.

[0062] The bias value adjustment formula is as follows:

[0063]

[0064] is the dynamically adjusted bias value, which is the EDAM mechanism for b k The bias value after dynamic adjustment; b k is the initial bias value, which is generated by random initialization (Gaussian distribution or uniform distribution) during the construction of the time series energy entropy network; μ is the bias threshold parameter, which is used to prevent numerical overflow and is set according to expert experience.

[0065] Finally, TEENet uses the comprehensive feature vector, the dynamically adjusted weight matrix and the dynamically adjusted bias value to predict the power load. The prediction formula is realized by nonlinearly combining the comprehensive feature vector, the dynamically adjusted weight matrix and the dynamically adjusted bias value, so that the predicted value has accuracy and stability. The prediction formula is as follows:

[0066]

[0067] in, is the predicted value at time point t+1, which is the prediction result of TEENet.

[0068] The predicted value is denormalized and restored to the scale of the original power load data. Subsequently, the denormalized predicted value is compared with the actual power load data, and the accuracy of the prediction result is evaluated by calculating the prediction error (such as mean square error or absolute percentage error). According to the specific application scenario and business needs, the error threshold is set in combination with expert experience. The prediction error is compared and analyzed with the set error threshold. If the prediction error is less than or equal to the error threshold, it means that the prediction accuracy of TEENet meets the requirements and can be directly used in practical applications (such as power dispatching, demand management, etc.). If the prediction error exceeds the error threshold, it means that the prediction accuracy of TEENet has not yet met the requirements, and adjustment and optimization operations are required to adjust the attenuation coefficient, bias threshold parameters and weight threshold parameters, and use the back propagation algorithm and gradient descent method to optimize the convolution kernel weight matrix and convolution bias vector to complete the adjustment and optimization of the time series energy entropy network.

[0069] After adjustment and optimization, use TEENet to forecast power load again, recalculate the forecast error and compare it with the set error threshold. If the forecast accuracy still does not meet the requirements, continue to adjust the parameters of the time series energy entropy network and iterate this process until the forecast error is less than or equal to the error threshold.

[0070] Finally, an analysis report of the forecast results is generated, including the forecast results and the evaluation of forecast accuracy. The analysis report provides decision makers with a scientific basis for decision-making and helps optimize decision-making.

[0071] In summary, a power load forecasting method based on deep learning was completed.

[0072] The order of the embodiments of the invention is for description only and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0073] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0074] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A power load forecasting method based on deep learning, characterized in that: The following steps are involved: S1: Collect and preprocess historical power load data to obtain power load data; construct a time series energy entropy network based on the power load data, calculate the energy value, and calculate the energy entropy based on the energy value; Based on the energy entropy, an energy entropy sequence matrix is ​​formed, and energy entropy features are extracted from the energy entropy sequence matrix; S2: Fusion of energy entropy features to generate a comprehensive feature vector; introduction of an energy entropy dynamic adjustment mechanism to dynamically adjust the weight matrix and bias term of the time series energy entropy network; prediction of the power load based on the comprehensive feature vector, the dynamically adjusted weight matrix, and the dynamically adjusted bias value; evaluation of the prediction results, optimization of the time series energy entropy network, use of the optimized time series energy entropy network to predict the power load, and generation of an analysis report of the prediction results.

2. The power load forecasting method based on deep learning according to claim 1, characterized in that: The S1 specifically includes: The time series energy entropy network consists of an input layer, a feature extraction layer, a feature fusion layer and an output layer.

3. The method for power load forecasting based on deep learning according to claim 2, characterized in that: The S1 specifically includes: The time series energy entropy network receives power load data through the input layer and calculates the energy value. The formula is as follows: Among them, E t is the energy value of the power load value at time point t; N is the number of power load values ​​at time point t; x t，i is the ith power load value at time point t; x t，j is the jth power load value at time point t; i is the power load value index; j is the index variable of the power load value summation operation; t is the time index, indicating the current time point; α is the attenuation coefficient; max(x t ) is the maximum value of all power load values ​​at time point t; is an exponential decay factor.

4. The power load forecasting method based on deep learning according to claim 1, characterized in that: The S1 specifically includes: The time scale and time window length are introduced. At each time scale, the energy entropy is arranged in chronological order to obtain an energy entropy sequence. Starting from the starting point of the energy entropy sequence, the time window is intercepted according to the time window length, and the time window slides backward, one time step at a time, until the entire energy entropy sequence is covered, and the energy entropy sequence matrix is ​​obtained. The energy entropy in each time window represents a row of the energy entropy sequence matrix.

5. The method for power load forecasting based on deep learning according to claim 4, characterized in that: The S1 specifically includes: The time series energy entropy network performs convolution operation through the feature extraction layer to extract energy entropy features from the energy entropy sequence matrix.

6. The method for power load forecasting based on deep learning according to claim 2, characterized in that: The S2 specifically includes: The time series energy entropy network fuses the energy entropy features through the feature fusion layer to form a comprehensive feature vector.

7. The method for power load forecasting based on deep learning according to claim 6, characterized in that: The S2 specifically includes: The output layer of the time series energy entropy network introduces an energy entropy dynamic adjustment mechanism. The energy entropy dynamic adjustment mechanism adjusts the weight matrix and bias value of the time series energy entropy network in real time according to the change of energy entropy to obtain a dynamically adjusted weight matrix and a dynamically adjusted bias value.

8. The method for power load forecasting based on deep learning according to claim 7, characterized in that: The S2 specifically includes: The power load is predicted using the comprehensive eigenvector, the dynamically adjusted weight matrix and the dynamically adjusted bias value; the prediction formula is as follows: in, is the predicted value of the time series energy entropy network at time point t+1; is the dynamically adjusted weight matrix; is the dynamically adjusted bias value; ⊙ is the element-by-element multiplication operation; F t is the comprehensive feature vector at time point t; ||·|| is the Euclidean length of the calculation vector.

9. The method for power load forecasting based on deep learning according to claim 8, characterized in that: The S2 specifically includes: The predicted value is denormalized and compared with the actual power load data. The prediction result is evaluated by calculating the prediction error. When the prediction error is greater than the preset error threshold, the time series energy entropy network is optimized and adjusted. The power load is predicted using the optimized time series energy entropy network until the prediction error is no greater than the error threshold, and an analysis report of the prediction result is generated.

Images