Photovoltaic prediction evolution deep learning model based on DCS differential creative search algorithm
Through the photovoltaic prediction deep learning model based on DCS differential creative search algorithm, combined with the encoder-decoder structure and multi-layer convolutional neural network, the prediction problem of photovoltaic power generation is solved, and high accuracy and real-time prediction is achieved As a result, it is suitable for planning and scheduling of power systems.
Patent Information
- Application Number
- CN202410949875.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-16
- Publication Date
- 2024-11-08
AI Technical Summary
The power of photovoltaic power generation is significantly volatile and intermittent due to factors such as solar radiation and ambient temperature, which leads to huge challenges in the planning, operation and scheduling of power systems, and it is difficult for existing technology to achieve accurate predictions.
The photovoltaic prediction evolution deep learning model based on DCS differential creative search algorithm is adopted, and the photovoltaic prediction evolution deep learning model is combined with data preprocessing, DCS differential creative search algorithm and encoder-decoder structure combined with multi-layer convolutional neural network, multi-head self-attention mechanism and long-term and short-term trends. The hybrid model is trained and the hyperparameters are automatically adjusted to improve prediction accuracy.
Accurate long-term prediction of photovoltaic power generation power is achieved, prediction errors are reduced, and the generalization ability and real-time nature of the model are improved. It can effectively capture data trends and process volatility data, meeting the planning and scheduling needs of the power system.
Smart Images

Figure CN118917684A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of data prediction, and in particular to a photovoltaic prediction evolutionary deep learning model based on a DCS differential creative search algorithm. Background Art
[0002] With the rapid growth of energy consumption and the continuous deterioration of the environment, it is particularly important to promote the sustainable development of human society. Therefore, new energy sources such as solar energy and wind energy have received great attention from the world. As a key technology, photovoltaic power generation has a rapid growth in installed capacity and has great development potential. However, the power of photovoltaic power generation is affected by factors such as solar radiation and ambient temperature, and has obvious volatility and intermittency. This has led to a large number of photovoltaic power stations (whether large-scale or distributed) connected to the power grid, which has brought huge challenges to the planning, operation, scheduling and control of the power system. Accurate prediction of photovoltaic power can not only provide a reliable basis for grid scheduling decisions and reduce the cost of electricity operation, but also provide strong technical support for the spatiotemporal complementary coordinated control of multiple energy sources. In this way, it can not only ensure the safe and stable operation of the power system, but also promote the optimized operation of the power grid, which has important practical significance. Summary of the invention
[0003] The purpose of the present invention is to provide a photovoltaic prediction evolutionary deep learning model. In view of the defects existing in the prior art, the purpose of the present invention is to achieve deep feature fusion and accurate prediction.
[0004] The technical solution adopted by the present invention is: a photovoltaic prediction evolutionary deep learning model based on DCS differential creative search algorithm;
[0005] A photovoltaic prediction evolutionary deep learning model based on a DCS differential creative search algorithm, characterized in that the described model has a data preprocessing part, a DCS differential creative search algorithm part, and a prediction part;
[0006] The data preprocessing part obtains weather data and photovoltaic power generation data from the outside, and standardizes and batch processes the externally obtained data.
[0007] The DCS differential creative search algorithm part inputs the required hyperparameters, sets the upper and lower limits, defines the fitness function, and then ranks the fitness of the obtained hyperparameters into two rankings, the top ranking and the bottom ranking. The hyperparameters of the top ranking are diverged and evolved, and the hyperparameters of the bottom ranking are refined within the upper and lower limits, and finally the hyperparameters with high fitness are obtained; the hyperparameters are: batch size, learning rate, number of multiple heads, and convolution kernel dimension.
[0008] The prediction module obtains one year of historical data and required prediction data from data preprocessing, obtains highly adaptive hyperparameters from the DCS differential creative search algorithm, uses an encoder-decoder structure, including a multi-layer convolutional neural network (CNN), a multi-head self-attention mechanism (MHSA), a combination of long-term and short-term trends (MIC) and their combinations for training, uses the trained prediction model to calculate one year of long-term predicted power generation data, inputs the predicted power generation data into the prediction results and evaluation index part for evaluation, and finally uses the used historical data and the obtained 1-year data to retrain the model. The historical data is a fusion of weather data and photovoltaic power generation data, and is characterized by: timestamp, active power delivery / received, current phase average, performance ratio, wind speed, weather temperature (degrees Celsius), weather relative humidity, global horizontal radiation, diffuse horizontal radiation, wind direction, daily rainfall, radiation global tilt surface, radiation diffuse tilt surface clear coefficient, scattered irradiance ratio, irradiance change trend, irradiance variance, active power;
[0009] The DCS differential creative search algorithm includes the following steps:
[0010] 1. Population
[0011] The population is randomly generated according to the required upper and lower limits of the hyperparameters. The population is a two-dimensional matrix with the number of rows being the number of populations and the number of columns being the number of required hyperparameters.
[0012] X i =X min +rand(0,1)×X max (1)
[0013] In the formula, X i is the i-th population; rand(0,1) is a random number between 0 and 1; X min and X max , are the lower and upper limits of the hyperparameters, respectively.
[0014] 2. Fitness Function
[0015]
[0016] In the formula, ILoss is the loss generated during model training; y is the actual photovoltaic power; To predict photovoltaic power.
[0017] 3. Divergence and Evolution
[0018]
[0019] In the formula, σ is the scale parameter, which affects the spread or dispersion of the distribution; α is the shape parameter, which affects the overall shape of the distribution; t is the input variable; S is a random symbol; Z represents the standard Laplace distribution, whose position is 0 and scale is 1; R represents the density of the positive random variable; X i+1 , for the next generation population; X i , for this generation population.
[0020] 4. Refining
[0021]
[0022] In the formula, NFE t , is the number of current function evaluations at time t; NFE max , the maximum number of function evaluations; X best : The individual with the best fitness value in the current generation; X r1 and X r2 : Two different individuals randomly selected from the population; W is the weight factor, which adjusts the optimal population X best The default value is 1; ω i,t , the tth time, the i-th random number uniformly distributed in the interval (0, 1).
[0023] The prediction module predicts the power generation including the following steps:
[0024] 1. Separation
[0025] Separate the timestamp from other features for subsequent step processing.
[0026] 2. Model Training
[0027] The prediction module uses the sampled values of the real-time weather data of similar days after normalization as the sample set of weather data, takes weather data as input and power generation as output, and trains a hybrid model using an encoder-decoder structure, including a multi-layer convolutional neural network (CNN), a multi-head self-attention mechanism (MHSA), a combination of long-term and short-term trends (MIC) and their combination, to obtain the weights between each layer;
[0028] 3. Prediction Results
[0029] The prediction module uses the numerical weather forecast data of the prediction day as the predicted weather data and the predicted weather data as input. It adopts the encoder-decoder structure and the trained multi-layer convolutional neural network (CNN), multi-head self-attention mechanism (MHSA), long-term and short-term trend combination (MIC) and their combined hybrid models to calculate and obtain the predicted power generation data and evaluate the data. Then, the predicted power generation data and the weather data used for prediction are added to the historical data used, and then the model training is updated again.
[0030] IV. Evaluation Indicators
[0031] 1. Mean Absolute Error (MAE)
[0032]
[0033] 2. Mean Squared Error (MSE)
[0034]
[0035] 3. Root Mean Squared Error (RMSE)
[0036]
[0037] 4. Mean Absolute Percentage Error (MAPE)
[0038]
[0039] 5. Mean Squared Percentage Error (MSPE)
[0040]
[0041] y i is the i-th true value; is the i-th predicted value; n is the number of samples;
[0042] The weather data includes timestamp, active power delivery / received, current phase average, performance ratio, wind speed, weather temperature (degrees Celsius), weather relative humidity, global horizontal radiation, diffuse horizontal radiation, wind direction, daily rainfall, radiation global tilt surface, radiation diffuse tilt surface clearness coefficient, diffuse irradiance ratio, irradiance change trend, and irradiance variance;
[0043] The hyperparameters are derived from the results of the DCS differential creative search algorithm, and the prediction part provides the DCS differential creative search algorithm part with a reporting module to obtain short-term predicted power generation data from the database every day.
[0044] The prediction model obtains weather data and photovoltaic power generation data from the outside as raw data, completes and corrects the raw sampled data, obtains sampled value data samples of weather data and photovoltaic power generation data respectively, and inputs the sampled value data samples into the photovoltaic prediction module.
[0045] The long-term prediction time range is the next 365 days, and the time resolution is 1 day of power generation prediction results. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for describing the embodiments are briefly introduced below.
[0047] Figure 1 It is a structural schematic diagram;
[0048] Figure 2 Schematic diagram of DCS differential creative search algorithm;
[0049] Figure 3 is a schematic diagram of an encoder-decoder;
[0050] Figure 4 Schematic diagram of a multi-layer convolutional neural network (CNN);
[0051] Figure 5 Schematic diagram of the multi-head self-attention mechanism (MHSA);
[0052] Figure 6 This is a schematic diagram of the combination of long-term and short-term trends (MIC);
[0053] Figure 7 Schematic diagram of training loss;
[0054] Figure 8 Schematic diagram of prediction results; DETAILED DESCRIPTION
[0055] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments.
[0056] The embodiments of the present invention are further described in detail below with reference to the accompanying drawings.
[0057] like Figure 1As shown, the photovoltaic prediction evolutionary deep learning model of the DCS differential creative search algorithm of the present invention has five parts, namely, a data preprocessing part, a prediction model part, a DCS algorithm part, a sliding window method part, a prediction result and an evaluation index part.
[0058] The data preprocessing part receives weather data and photovoltaic power generation data from the outside and standardizes the data. The input format of weather data and photovoltaic power generation data consists of fields, field types and descriptions, as shown in Table 1.
[0059] Table 1 Input format of weather data and photovoltaic power generation data
[0060]
[0061]
[0062] The prediction module performs long-term prediction, that is, inputs 365 days of data and predicts 365 days of data. The sliding window method adds the predicted data to the historical data to update the model.
[0063] The prediction module receives the data standardized by the data preprocessing module, namely the weather data and photovoltaic power generation data, completes and corrects them, and processes the timestamp features and other features separately to obtain sampling value data samples of the timestamp data and other feature data respectively.
[0064] The DCS differential creative search algorithm includes the following steps: Figure 2 As shown:
[0065] 1. Population
[0066] The population is randomly generated according to the required upper and lower limits of the hyperparameters. The population is a two-dimensional matrix with the number of rows being the number of populations and the number of columns being the number of required hyperparameters.
[0067] X i =X min +rand(0,1)×X max (10)
[0068] In the formula, X i is the i-th population; rand(0,1) is a random number between 0 and 1; X min and X max , are the lower and upper limits of the hyperparameters, respectively.
[0069] 2. Fitness Function
[0070]
[0071] In the formula, ILoss is the loss generated during model training; y is the actual photovoltaic power; To predict photovoltaic power.
[0072] 3. Divergence and Evolution
[0073]
[0074] In the formula, σ is the scale parameter, which affects the spread or dispersion of the distribution; α is the shape parameter, which affects the overall shape of the distribution; t is the input variable; S is a random symbol; Z represents the standard Laplace distribution, whose position is 0 and scale is 1; R represents the density of the positive random variable; X i+1 , for the next generation population; X i , for this generation population.
[0075] 4. Refining
[0076]
[0077] In the formula, NFE t , is the number of current function evaluations at time t; NFE max , the maximum number of function evaluations; X best : The individual with the best fitness value in the current generation; X r1 and X r2 : Two different individuals randomly selected from the population; W is the weight factor, which adjusts the optimal population X best The default value is 1; ω i,t , the tth time, the i-th random number uniformly distributed in the interval (0, 1).
[0078] The prediction module predicts the power generation including the following steps:
[0079] 1. Separation
[0080] Separate the timestamp from other features for subsequent step processing.
[0081] 2. Model Training
[0082] The prediction module uses the sampled values of the real-time weather data of similar days and the values after normalization as the sample set of weather data, takes weather data as input and power generation as output, and adopts an encoder-decoder structure, such as Figure 3 , including multi-layer convolutional neural networks (CNNs), such as Figure 4 , Multi-head Self-Attention Mechanism (MHSA), such as Figure 5 , long-term and short-term trend combination (MIC), such as Figure 6 and its combined hybrid model to obtain the weights between layers; training loss, such as Figure 7 ;
[0083] 3. Prediction Results
[0084] The prediction module uses the numerical weather forecast data of the prediction day as the predicted weather data and the predicted weather data as input. It adopts the encoder-decoder structure and the trained multi-layer convolutional neural network (CNN), multi-head self-attention mechanism (MHSA), long-term and short-term trend combination (MIC) and their combined hybrid models to calculate and obtain the predicted power generation data and evaluate the data. Then, the predicted power generation data and the weather data used for prediction are added to the historical data used, and then the model training is updated again.
[0085] The system of the present invention predicts the photovoltaic power generation power, and takes the data prediction results from January 1, 2012 to January 1, 2013 as an example for comparison. A curve is drawn with time as the horizontal axis and power as the vertical axis, as shown in FIG. Figure 8 As shown, by comparing the degree of agreement between the predicted power curve and the actual power curve, it can be seen that the predicted and actual power curves basically coincide with each other, indicating that the photovoltaic power predicted by the system of the present invention has a relatively small error and is accurate, practical and effective.
[0086] IV. Evaluation Indicators
[0087] 1. Mean Absolute Error (MAE)
[0088]
[0089] 2. Mean Squared Error (MSE)
[0090]
[0091] 3. Root Mean Squared Error (RMSE)
[0092]
[0093] 4. Mean Absolute Percentage Error (MAPE)
[0094]
[0095] 5. Mean Squared Percentage Error (MSPE)
[0096]
[0097] y i is the i-th true value; is the i-th predicted value; n is the number of samples;
[0098] The post-prediction evaluation indicators are shown in Table 2.
[0099] Table 2 Evaluation indicators
[0100]
[0101]
[0102] From Table 2 and Figure 8 It can be seen that in the one-year forecast process, for some data points with large fluctuations, the model has a good degree of fit and high prediction accuracy, which meets the basic requirements of power system planning and dispatching. This shows that the forecasting model performs well in processing large fluctuations in data, not only capturing the main trends in the data, but also making relatively accurate predictions of abnormal fluctuations.
[0103] Through the analysis of the embodiments, it can be found that the photovoltaic prediction evolution deep learning model based on the DCS differential creative search algorithm can further refine the regularity of features, deeply mine features and learn from them, and is more sensitive to power fluctuations, thereby further improving the accuracy of the mapping relationship and being able to relatively effectively solve the problem of violent fluctuations in photovoltaic power prediction. From the overall error analysis results, the photovoltaic power prediction system of the present invention has achieved good prediction indicators.
[0104] The system of the present invention realizes a photovoltaic prediction evolution deep learning model based on a DCS differential creative search algorithm, automatically obtains hyperparameters through the DCS differential creative search algorithm, and is easy to debug the model. The system of the present invention establishes a deep learning model, adopts an encoder-decoder structure and a hybrid model of a trained multi-layer convolutional neural network (CNN), a multi-head self-attention mechanism (MHSA), a combination of long-term and short-term trends (MIC) and their combinations, uses nonlinear mapping and generalization capabilities to train the model, and combines weather forecast data to accurately predict the power generation.
[0105] In the present invention, the sliding window method is used to effectively process time series data. In particular, in photovoltaic prediction, the sliding window of historical data can be used to gradually predict future photovoltaic power generation. Through the sliding window method, when training and evaluating the prediction model, data from multiple time windows can be used for training and verification, thereby improving the generalization ability and accuracy of the model. The sliding window method allows the model to be dynamically updated. With the addition of new data, the prediction model can be continuously optimized and adjusted to maintain its sensitivity and adaptability to the data. The application of the sliding window method in photovoltaic prediction can effectively improve the prediction accuracy and the real-time performance of the model, making it more adaptable to the dynamically changing photovoltaic power generation environment.
[0106] The present invention not only performs well overall, but can also maintain high-precision predictions for data points with large fluctuations, fully meeting the needs of power system planning and scheduling, and contributing to the efficient and stable operation of the power system.
Claims
1. A photovoltaic prediction evolutionary deep learning model based on DCS differential creative search algorithm, Features: The model consists of a data preprocessing part, a DCS differential creative search algorithm part, and a prediction part; The data preprocessing part obtains weather data and photovoltaic power generation data from the outside, and standardizes and batch processes the externally obtained data; The DCS differential creative search algorithm inputs the required hyperparameters, sets the upper and lower limits, defines the fitness function, and then ranks the fitness of the obtained hyperparameters into two rankings, the top ranking and the bottom ranking. The hyperparameters of the top ranking are diverged and evolved, and the hyperparameters of the bottom ranking are refined within the upper and lower limits, and finally the hyperparameters with high fitness are obtained; the hyperparameters are: batch size, learning rate, number of multi-heads, and convolution kernel dimension; The prediction module obtains one year of historical data and required prediction data from data preprocessing, obtains highly adaptive hyperparameters from the DCS differential creative search algorithm, adopts an encoder-decoder structure, including a multi-layer convolutional neural network, a multi-head self-attention mechanism, a combination of long-term and short-term trends, and a combination thereof for training, uses the trained prediction model to calculate one year of long-term predicted power generation data, inputs the predicted power generation data into the prediction result and evaluation index part for evaluation, and finally uses the used historical data and the obtained 1-year data to retrain the model, the historical data is a fusion of weather data and photovoltaic power generation data, and is characterized by: timestamp, active power delivery / receipt, current phase average, performance ratio, wind speed, weather temperature, weather relative humidity, global horizontal radiation, diffuse horizontal radiation, wind direction, daily rainfall, radiation global tilt surface, radiation diffuse tilt surface clear coefficient, scattered irradiance ratio, irradiance change trend, irradiance variance, active power; Wherein, the DCS differential creative search algorithm comprises the following steps:
1. Randomly generate a population according to the required upper and lower limits of the hyperparameters. The population is a two-dimensional matrix with the number of rows being the number of populations and the number of columns being the number of required hyperparameters. The specific calculation method is: X i =X min +rand(0,1)×X max In the formula, X i is the i-th population; rand(0,1) is a random number between 0 and 1; X min and X max , are the lower and upper limits of the hyperparameters, respectively.
2. Fitness Function In the formula, ILoss is the loss generated during model training; y is the actual photovoltaic power; To predict photovoltaic power.
3. Divergence and Evolution In the formula, σ is the scale parameter, which affects the spread or dispersion of the distribution; α is the shape parameter, which affects the overall shape of the distribution; t is the input variable; S is a random symbol; Z represents the standard Laplace distribution, whose position is 0 and scale is 1; R represents the density of the positive random variable; X i+1 , for the next generation population; X i , for this generation population.
4. Refining In the formula, NFE t , is the number of current function evaluations at time t; NFE max , Maximum number of function evaluations; X best : The individual with the best fitness value in the current generation; X r1 and X r2 : Two different individuals randomly selected from the population; W is the weight factor, which adjusts the optimal population X best The default value is 1; ω i,t , the tth time, the i-th random number uniformly distributed in the interval (0, 1); The prediction module predicts the power generation including the following steps:
1. Separation Separate the timestamp from other features for subsequent step processing.
2. Model Training The prediction module uses the sampled values of the real-time weather data of similar days after normalization as the sample set of weather data, takes weather data as input and power generation as output, and trains a hybrid model using an encoder-decoder structure, including a multi-layer convolutional neural network (CNN), a multi-head self-attention mechanism (MHSA), a combination of long-term and short-term trends (MIC) and their combination, to obtain the weights between each layer; 3. Prediction Results The prediction module uses the numerical weather forecast data of the prediction day as the predicted weather data and the predicted weather data as input. It adopts the trained multi-layer convolutional neural network (CNN), multi-head self-attention mechanism (MHSA), long-term and short-term trend combination (MIC) and their combined hybrid models to calculate and obtain the predicted power generation data and evaluate the data. Then, the predicted power generation data and the weather data used for prediction are added to the historical data used, and then the model training is updated again.
2. The photovoltaic power prediction model according to claim 1, Features: The weather data includes timestamp, active power delivery / received, current phase average, performance ratio, wind speed, weather temperature (degrees Celsius), weather relative humidity, global horizontal radiation, diffuse horizontal radiation, wind direction, daily rainfall, radiation global tilt surface, radiation diffuse tilt surface clearness coefficient, scattered irradiance ratio, irradiance change trend, and irradiance variance.
3. The photovoltaic power prediction model according to claim 2, Features: The hyperparameters are derived from the results of the DCS differential creative search algorithm, and the prediction part provides the DCS differential creative search algorithm part with a reporting module to obtain short-term predicted power generation data from the database every day.
4. The photovoltaic power prediction model according to claim 3, Features: The prediction model obtains weather data and photovoltaic power generation data from the outside as raw data, completes and corrects the raw sampled data, obtains sampled value data samples of weather data and photovoltaic power generation data respectively, and inputs the sampled value data samples into the photovoltaic prediction module.
5. The photovoltaic power prediction model according to claim 4, Features: The long-term prediction time range is the next 365 days, and the time resolution is 1 day of power generation prediction results.
Citation Information
Cited By
Cloud-based industrial park micro-grid regulation and control system and method
CN120073765A
