A time-period-multiple-task learning photovoltaic power day-ahead prediction method and system
By employing a time-segmented, multi-task learning approach, and combining Bi-LSTM, WD-Bi-LSTM, and CNN models, the overfitting problem of deep learning in photovoltaic power generation prediction is solved. This achieves highly adaptable and refined photovoltaic power prediction under extreme weather conditions, supporting the effectiveness of grid dispatch.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANDONG UNIV
- Filing Date
- 2022-08-17
- Publication Date
- 2026-05-15
AI Technical Summary
Deep learning suffers from overfitting in photovoltaic power generation forecasting, resulting in poor adaptability under complex weather conditions. Furthermore, on-site mapping strategies based on numerical weather forecasts cannot accurately characterize power fluctuations, affecting the guidance effect of grid dispatching.
A time-segmented, multi-task learning approach is adopted, combining deep learning and multi-task learning frameworks. By integrating historical similarity, adjacency, and meteorological data through Bi-LSTM, WD-Bi-LSTM, and CNN models, the photovoltaic power trend is predicted. Genetic algorithms are used to optimize the weights of each time period to achieve refined prediction.
Under extreme weather conditions, it achieves high adaptability and fine-grained fluctuation characterization of photovoltaic power forecasting, improving the accuracy and stability of forecasting and supporting the effectiveness of grid dispatching.
Smart Images
Figure CN115409249B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power prediction technology, and relates to a time-segmented, multi-task learning-based method and system for predicting photovoltaic power day-ahead. Background Technology
[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.
[0003] Against the backdrop of "carbon peaking and carbon neutrality," photovoltaic (PV) power generation has become a mainstay of energy transition. However, as the proportion of PV power generation in the power grid's power structure increases, the randomness and intermittency of PV power generation pose significant challenges to the grid's development of reliable dispatch plans. Accurate and adaptable PV power generation forecasting has become a key technology for solving this problem.
[0004] Deep learning, with its advantages in nonlinear processing and feature extraction, is widely used in photovoltaic power prediction research. However, it often falls into the problem of overfitting under complex weather conditions, resulting in poor adaptability. Conversely, the mapping strategy based on numerical weather prediction (NWP) in the field abandons the detailed prediction of fluctuations under complex weather conditions. Although it ensures the adaptability and stability of the prediction model, the missing fluctuation description has limited guidance for grid dispatch. Summary of the Invention
[0005] To address the aforementioned issues, this invention proposes a time-segmented, multi-task learning-based method and system for predicting photovoltaic power day-ahead. Based on deep learning and a multi-task learning-time-segmented integration framework, this invention makes predictions more targeted and adaptable, and can maintain stability while depicting power fluctuation details even under extreme weather conditions.
[0006] According to some embodiments, the present invention adopts the following technical solution:
[0007] A time-segmented, multi-task learning-based method for predicting photovoltaic power day-ahead includes the following steps:
[0008] Based on the predicted values of irradiance and temperature, the approximate power trend is predicted;
[0009] Analyze the characteristics of adjacent days to predict power trends with adjacent dates;
[0010] Extract historical similarities under the same weather type to predict power trends with similarity;
[0011] Based on the general power trend, adjacent and similar power trends, a genetic algorithm is used to solve for the optimal ratio of each power trend. After integrating the power trends by time-segmentation integration, the final prediction result is obtained.
[0012] As an alternative implementation, the prediction of approximate power trends, adjacent and similar power trends is achieved through a hybrid deep learning model. This hybrid deep learning model maps the relationship between future power and relevant input factors for prediction. The relevant input factors include historical power information, meteorological measurement data, irradiance, and temperature forecast values.
[0013] As a further limitation, the overall timeline of the hybrid deep learning model includes the union of historical time steps and future time steps.
[0014] As a further limitation, the hybrid deep learning model includes parallel mapping models, temporal models, and convolutional neural network models.
[0015] As a further limitation, the mapping model is a Bi-LSTM model, and the time series model is a WD-Bi-LSTM model.
[0016] The mapping model determines the mapping relationship between measured meteorological data and historical power data;
[0017] The time series model uses WD to decompose the historical power sequence of the three nearest adjacent days into a main low-frequency sequence and two secondary high-frequency series according to different frequencies, which respectively represent the general trend and detailed fluctuation of photovoltaic power under different weather types. Combined with different LSTM models, it captures the nonlinear relationship and deep coupling characteristics of adjacent days under different frequencies, and generates a power trend with adjacent characteristics.
[0018] The convolutional neural network model extracts historical similar features under the same weather type, considers the photovoltaic power generation pattern under different weather conditions, and divides the clusters corresponding to the historical photovoltaic power generation data into several weather conditions.
[0019] Based on the weather type of the predicted day, similar historical weather data are clustered and passed to different convolutional neural network models to obtain power results with historical similarity characteristics.
[0020] As an alternative implementation method, the specific process of integrating various power trends by time-segmentation integration is as follows: taking into account the impact of weather conditions changing at different times of the day on photovoltaic output, the principle that the power trend ramp-up process is less affected by meteorological factors and the intermediate period is more affected by meteorological factors, the weights of each prediction result are allocated.
[0021] A time-segmented, multi-task learning-based photovoltaic power day-ahead prediction system includes:
[0022] The mapping module is configured to predict approximate power trends based on irradiance and temperature forecasts.
[0023] The time series extraction module is configured to predict power trends with adjacent days based on the features of adjacent days;
[0024] The similarity prediction module is configured to extract historical similarity features under the same weather type and predict power trends with similarity.
[0025] The time-segmented integration module is configured to use a genetic algorithm to determine the optimal ratio of each power trend based on the approximate power trend, adjacent and similar power trends, and then integrate the various power trends through the time-segmented integration method to obtain the final prediction result.
[0026] A computer-readable storage medium storing a plurality of instructions adapted for loading by a processor of a terminal device and executing steps in the method.
[0027] A terminal device includes a processor and a computer-readable storage medium, the processor being configured to implement instructions; the computer-readable storage medium being configured to store a plurality of instructions adapted to be loaded by the processor and executed in accordance with the steps of the method described therein.
[0028] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0029] To overcome the overfitting problem of deep learning when weather types change, and to balance the strong adaptability and accurate prediction of fluctuation details, this invention establishes a reasonable multi-task learning architecture to fully leverage the guiding role of NWP information in predicting future photovoltaic power. By using a hybrid deep learning method to mine historical similarities and adjacencies, and combining it with a time-segmented integration method, the prediction architecture becomes more targeted and adaptable, and can also characterize the details of power fluctuations under extreme weather conditions. Attached Figure Description
[0030] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0031] Figure 1 This is a schematic diagram of the photovoltaic power day-ahead forecasting method.
[0032] Figure 2 This is a flowchart illustrating the hybrid deep learning approach.
[0033] Figure 3 This is a flowchart illustrating the time-segmented integration method. Detailed implementation method:
[0034] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0035] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, 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 pertains.
[0036] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0037] like Figure 1 As shown, a photovoltaic power day-ahead prediction method based on time-segmented multi-task learning includes the following steps:
[0038] Based on the predicted values of irradiance and temperature, the approximate power trend is predicted;
[0039] Based on the characteristics of adjacent days, predict the power trend with adjacency;
[0040] Extract historical similarities under the same weather type to predict power trends with similarity;
[0041] Based on the general power trend, adjacent and similar power trends, a genetic algorithm is used to solve for the optimal ratio of each power trend. After integrating the power trends by time-segmentation integration, the final prediction result is obtained.
[0042] Among them, the approximate power trend, the adjacent power trend, and the similarity power trend are achieved through a hybrid deep learning model.
[0043] like Figure 2 As shown, the hybrid deep learning model makes predictions by mapping the relationship between future power and relevant input factors. The overall time axis is defined as historical time step T. h and future time step T f The union of T h ∪T f ={t1,t2,…,t h}∪{t h+1 ,t h+2 ,…,t h+f Based on historical power information Meteorological measurement data NWP output The goal is to generate the final prediction result using model weights W.
[0044]
[0045]
[0046] Represents the prediction output matrix of different models, including and W represents the prediction output of the Bi-LSTM mapping, WD-Bi-LSTM temporal model, and CNN model, respectively. MAP W TSM and W CNN These represent the hyperparameters of their respective models. The function f(·) can be represented using a neural network architecture.
[0047] Considering the influence of external meteorological patterns and the differences in internal photovoltaic power generation status during different time periods, a time-segmented integration method is used to analyze the coefficient matrix λ for different time periods. i The optimization is performed using (m*n), as shown in equation (3). m is the total number of time periods, and n is the number of prediction models.
[0048]
[0049] Specifically, to address the overfitting problem of deep learning methods in situations of frequent weather changes, a Bi-LSTM-based mapping model was established using NWP data from a third-party institution. This model predicted a highly adaptive power trend in the first task. However, due to the uncertainty of historical day-ahead NWP information, there are many incorrect mapping relationships between historical NWP and historical power. Directly establishing a mapping model would inevitably lead to strong overfitting. Therefore, a mapping model should be trained based on measured meteorological data and historical power data.
[0050] To compensate for the inability of the above mapping models to capture the adjacent characteristics of historical photovoltaic power and to overcome the strong fluctuations in photovoltaic power output at a 15-minute resolution, a WD-Bi-LSTM time series model is used to analyze the characteristics of adjacent days. Unlike the optimized mapping model based on meteorological information, the WD-Bi-LSTM time series model is used to analyze and extract historical adjacent characteristics: using WD, the historical power series of the last three adjacent days are decomposed into a main low-frequency series and two secondary high-frequency series according to different frequencies, which respectively represent the general trend and detailed fluctuations of photovoltaic power under different weather types. Combined with different LSTM models, the nonlinear relationships and deep coupling characteristics of adjacent days at different frequencies are captured, thereby generating a power trend with adjacent characteristics.
[0051] In addition to adjacent features, a CNN model is used to extract historical similarity features under the same weather type. Considering the photovoltaic power generation patterns under different weather conditions (sunny, cloudy, rainy, overcast), the historical photovoltaic power generation data is first divided into clusters corresponding to four weather conditions. Four types of CNN models are then established according to different weather types to extract historical similarity characteristics of power output under different weather types. Furthermore, based on the weather type of the next day, historical power information under the corresponding weather type is selected and passed to the trained CNN model to generate power results with historical similarity characteristics.
[0052] like Figure 3 As shown, considering the varying impacts of weather conditions and unit output characteristics on photovoltaic (PV) output at different times of the day, the PV output within a day is divided into three phases based on the cumulative correlation analysis of historical meteorological and power information: output ramp-up phase, fluctuation phase, and output deceleration phase. The two phases of the daily PV output show strong historical adjacency, while the fluctuation phase in the middle is significantly affected by meteorological factors. Therefore, during model learning, to quantify the differences in the impact of external meteorological factors and internal module characteristics on PV output at different times, the approximate power trend provided by the NWP mapping model, the adjacency characteristics extracted by WD-Bi-LSTM, and the historical similarity characteristics mined by CNN are finely integrated according to the integration ratio for different time periods. This ensures model adaptability while depicting the details of PV output fluctuations. The integration ratio of each model output in each time period is calculated using the GA algorithm, with the optimization objective being to minimize the error between the power measurement value within the past week and the predicted power value before the day. Finally, the ratios are summed to obtain the final power output.
[0053] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0054] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0055] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0056] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0057] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
[0058] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A time-segmented, multi-task learning-based method for predicting photovoltaic power day-ahead, characterized in that, Includes the following steps: Based on the predicted values of irradiance and temperature, the approximate power trend is predicted; Based on the characteristics of adjacent days, predict the power trend with adjacency; Extract historical similarities under the same weather type to predict power trends with similarity; Based on the general power trend, adjacent and similar power trends, a genetic algorithm is used to solve for the optimal ratio of each power trend. After integrating the power trends by time-segmentation integration, the final prediction result is obtained. The prediction of approximate power trends, adjacent and similar power trends is achieved through a hybrid deep learning model. The hybrid deep learning model maps the relationship between future power and relevant input factors for prediction. The relevant input factors include historical power information, meteorological measurement data and irradiance and temperature forecast values. The hybrid deep learning model includes parallel mapping models, temporal models, and convolutional neural network models; The mapping model is a Bi-LSTM model, and the time series model is a WD-Bi-LSTM model; The mapping model determines the mapping relationship between measured meteorological data and historical power data; The time series model uses WD to decompose the historical power sequence of the three nearest adjacent days into a main low-frequency sequence and two secondary high-frequency series according to different frequencies, which respectively represent the general trend and detailed fluctuation of photovoltaic power under different weather types. Combined with different LSTM models, it captures the nonlinear relationship and deep coupling characteristics of adjacent days under different frequencies, and generates a power trend with adjacent characteristics. The convolutional neural network model extracts historical similar features under the same weather type, considers the photovoltaic power generation pattern under different weather conditions, and divides the clusters corresponding to the historical photovoltaic power generation data into several weather conditions. Based on the weather type of the predicted day, similar historical weather data are clustered and passed to different convolutional neural network models to obtain power results with historical similarity characteristics.
2. The photovoltaic power day-ahead prediction method based on time-segmented, multi-task learning as described in claim 1, characterized in that, The overall timeline of the hybrid deep learning model includes the union of historical time steps and future time steps.
3. The photovoltaic power day-ahead prediction method based on time-segmented, multi-task learning as described in claim 1, characterized in that, The specific process of integrating various power trends using the time-segmented integration method is as follows: considering the impact of weather conditions changing at different times of the day on photovoltaic output, and based on the principle that the power trend ramp-up process is less affected by meteorological factors and the intermediate period is more affected by meteorological factors, the weights of each prediction result are allocated.
4. A photovoltaic power day-ahead prediction system based on time-segmented, multi-task learning, employing the method described in claim 1, characterized in that, include: The mapping module is configured to predict approximate power trends based on irradiance and temperature forecasts. The time series extraction module is configured to predict power trends with adjacent days based on the features of adjacent days; The similarity prediction module is configured to extract historical similarity features under the same weather type and predict power trends with similarity. The time-segmented integration module is configured to use a genetic algorithm to determine the optimal ratio of each power trend based on the approximate power trend, adjacent and similar power trends, and then integrate the various power trends through the time-segmented integration method to obtain the final prediction result.
5. A computer-readable storage medium, characterized in that, It stores multiple instructions, which are adapted to be loaded by the processor of the terminal device and executed as steps in the method of any one of claims 1-3.
6. A terminal device, characterized in that, It includes a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store a plurality of instructions adapted to be loaded by the processor and executed as steps in the method of any one of claims 1-3.