Method, device and method for training photovoltaic output prediction model of distributed photovoltaic system based on minute-level data

Abstract

This application relates to a training method, apparatus, and prediction method for a photovoltaic (PV) output prediction model of a distributed PV system based on minute-level data. The method includes acquiring PV output-related data of the distributed PV system; preprocessing the PV output-related data to obtain preprocessed data; extracting features from the preprocessed data to obtain time-series feature data, meteorological feature data, and spatial feature data; and training the original dynamic recurrent neural network model using the time-series feature data, meteorological feature data, and spatial feature data to obtain the target prediction model. This method can improve prediction accuracy.

Description

Technical Field

[0001] This application relates to the field of photovoltaic power output prediction technology for distributed photovoltaic systems, and in particular to a training method, apparatus and prediction method for a photovoltaic power output prediction model for distributed photovoltaic systems based on minute-level data. Background Technology

[0002] As the proportion of distributed photovoltaic power generation in the power system continues to increase, accurate power forecasting is becoming increasingly crucial for the safe and stable operation and efficient dispatch of the power grid.

[0003] However, existing technologies have significant shortcomings. For example, traditional methods suffer from insufficient data timeliness, which limits prediction accuracy. Conventional models such as ARIMA and support vector machines are difficult to adapt to the strong volatility and spatial heterogeneity of photovoltaic power output, resulting in poor modeling performance. Furthermore, the predictions only focus on the photovoltaic power generation load itself, lacking key information such as the inverter power generation ratio, the grid load on the distribution network, and the user-side absorption load. This makes it impossible to provide comprehensive support for grid dispatch and restricts the efficient utilization of distributed photovoltaic power generation. Summary of the Invention

[0004] Therefore, it is necessary to provide a training method, device, and prediction method for a photovoltaic power output prediction model of a distributed photovoltaic system based on minute-level data, addressing the aforementioned technical problems.

[0005] Firstly, this application provides a method for training a photovoltaic output prediction model for a distributed photovoltaic system based on minute-level data, the method comprising:

[0006] Acquire photovoltaic output data related to distributed photovoltaic systems; the photovoltaic output data includes minute-by-minute power load data and minute-by-minute meteorological data for distributed photovoltaic power stations located in different geographical locations within a preset historical time period;

[0007] The photovoltaic power output data is preprocessed to obtain the preprocessed data.

[0008] Feature extraction is performed on the preprocessed data to obtain time-series feature data, meteorological feature data, and spatial feature data. The time-series feature data represents the trend and fluctuation of the power load data per minute within a preset historical time period. The meteorological feature data represents the lag of the power load data per minute relative to the meteorological data per minute within a preset historical time period. The spatial feature data represents the impact of the geographical location of the distributed photovoltaic power station on the power load data per minute within a preset historical time period.

[0009] The original dynamic recurrent neural network model was trained using time-series feature data, meteorological feature data, and spatial feature data to obtain the target prediction model.

[0010] In one embodiment, the original dynamic recurrent neural network model is trained using time-series feature data, meteorological feature data, and spatial feature data to obtain a target prediction model, including:

[0011] By substituting time-series feature data, meteorological feature data, and spatial feature data into the original dynamic recurrent neural network model, and adjusting the parameters of the original dynamic recurrent neural network model through an adaptive momentum estimation algorithm and a learning rate decay strategy, the performance index of the original dynamic recurrent neural network model reaches the preset conditions, thereby obtaining the target prediction model.

[0012] In one embodiment, the performance metric includes the target loss function value, the minimum value of which corresponds to a preset condition. The target loss function is a composite function consisting of the mean squared error function and the mean absolute error function. Time-series feature data, meteorological feature data, and spatial feature data are substituted into the original dynamic recurrent neural network model. The parameters of the original dynamic recurrent neural network model are adjusted using an adaptive momentum estimation algorithm and a learning rate decay strategy to ensure that the performance metric of the original dynamic recurrent neural network model meets the preset condition, thereby obtaining the target prediction model, including:

[0013] The minimum value of the target loss function of the original dynamic recurrent neural network model is calculated using an adaptive momentum estimation algorithm and a learning rate decay strategy.

[0014] The parameters of the original dynamic recurrent neural network model are adjusted based on the minimum value of the target loss function to obtain the target prediction model.

[0015] In one embodiment, the photovoltaic output data of the distributed photovoltaic system is preprocessed to obtain preprocessed data, including:

[0016] Outliers in photovoltaic power output data were removed using a sliding window method.

[0017] The remaining photovoltaic output data after removing outliers are normalized to obtain preprocessed data.

[0018] In one embodiment, prior to the step of training the original dynamic recurrent neural network model using time-series feature data, meteorological feature data, and spatial feature data, the method further includes:

[0019] Clustering time-series characteristic data, meteorological characteristic data, and spatial characteristic data belonging to the same time window yields multiple clusters;

[0020] Multiple clusters are used as training sample sets and substituted into the original dynamic recurrent neural network model.

[0021] Secondly, this application provides a method for predicting the photovoltaic output of a distributed photovoltaic system, the method comprising:

[0022] Acquire photovoltaic output data related to distributed photovoltaic systems. This data includes minute-by-minute power load data and minute-by-minute meteorological data for distributed photovoltaic power stations located in different geographical locations within a preset historical time period. The preset historical time period is a period prior to the time period to be predicted.

[0023] The photovoltaic power output data is preprocessed to obtain the preprocessed data.

[0024] Feature extraction is performed on the preprocessed data to obtain time-series feature data, meteorological feature data, and spatial feature data. The time-series feature data represents the trend and fluctuation of the power load data per minute within a preset historical time period. The meteorological feature data represents the lag of the power load data per minute relative to the meteorological data per minute within a preset historical time period. The spatial feature data represents the impact of the geographical location of the distributed photovoltaic power station on the power load data per minute within a preset historical time period.

[0025] The time-series feature data, meteorological feature data, and spatial feature data are input into the target prediction model to obtain the photovoltaic power output within the time period to be predicted. The target prediction model is trained using the training method of the photovoltaic power output prediction model of the distributed photovoltaic system based on minute-level data in the above embodiment.

[0026] Thirdly, this application also provides a training device for a photovoltaic output prediction model of a distributed photovoltaic system based on minute-level data, the device comprising:

[0027] The data acquisition module is used to acquire photovoltaic output-related data of distributed photovoltaic systems. The photovoltaic output-related data includes minute-by-minute power load data and minute-by-minute meteorological data for distributed photovoltaic power stations located in different geographical locations within a preset historical time period.

[0028] The preprocessing module is used to preprocess photovoltaic output-related data to obtain preprocessed data;

[0029] The feature extraction module is used to extract features from the preprocessed data to obtain time-series feature data, meteorological feature data, and spatial feature data. The time-series feature data represents the trend and fluctuation of the power load data per minute within a preset historical time period. The meteorological feature data represents the lag of the power load data per minute relative to the meteorological data per minute within a preset historical time period. The spatial feature data represents the impact of the geographical location of the distributed photovoltaic power station on the power load data per minute within a preset historical time period.

[0030] The training module is used to train the original dynamic recurrent neural network model using time-series feature data, meteorological feature data, and spatial feature data to obtain the target prediction model.

[0031] Fourthly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the methods described above.

[0032] Fifthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the methods described above.

[0033] Sixthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the methods described above.

[0034] The training method, apparatus, and prediction method of the photovoltaic output prediction model for distributed photovoltaic systems based on minute-level data described above have at least the following beneficial effects:

[0035] Based on 1-minute high-frequency power load and meteorological data, this method captures minute-level fluctuations in photovoltaic output, addressing the problem of insufficient timeliness in traditional data. By preprocessing to ensure data quality, it extracts three types of features—time series, meteorology, and space—to adapt to the strong volatility and spatial heterogeneity of photovoltaics, overcoming the poor adaptability of models and significantly improving the accuracy and real-time performance of ultra-short-term forecasts. This provides comprehensive information for grid dispatch, helping to optimize resource allocation, improve the stability of new energy grid connection, and enhance the efficiency of green energy consumption. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is an application environment diagram of a training method for a photovoltaic power output prediction model of a distributed photovoltaic system based on minute-level data in one embodiment;

[0038] Figure 2 This is a flowchart illustrating the training method for a photovoltaic output prediction model of a distributed photovoltaic system based on minute-level data in one embodiment.

[0039] Figure 3In one embodiment, a flowchart illustrates the steps of substituting time-series feature data, meteorological feature data, and spatial feature data into the original dynamic recurrent neural network model, adjusting the parameters of the original dynamic recurrent neural network model through an adaptive momentum estimation algorithm and a learning rate decay strategy, so that the performance index of the original dynamic recurrent neural network model reaches the preset conditions, thereby obtaining the target prediction model.

[0040] Figure 4 This is a flowchart illustrating the steps of preprocessing photovoltaic power output data of a distributed photovoltaic system to obtain preprocessed data in one embodiment.

[0041] Figure 5 This is a flowchart illustrating the training method for a photovoltaic output prediction model of a distributed photovoltaic system based on minute-level data in another embodiment.

[0042] Figure 6 This is a flowchart illustrating a method for predicting the photovoltaic output of a distributed photovoltaic system in one embodiment.

[0043] Figure 7 This is a structural block diagram of a training device for a photovoltaic power output prediction model of a distributed photovoltaic system based on minute-level data in one embodiment.

[0044] Figure 8 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0046] The training method for predicting photovoltaic output of a distributed photovoltaic system based on minute-level data provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed on a cloud or other network server. Server 104 acquires photovoltaic output-related data from the distributed photovoltaic system; this data includes minute-by-minute power load data and minute-by-minute meteorological data for distributed photovoltaic power stations located in different geographical locations within a preset historical time period; preprocesses the photovoltaic output-related data to obtain preprocessed data; extracts features from the preprocessed data to obtain time-series feature data, meteorological feature data, and spatial feature data; the time-series feature data characterizes the trend and fluctuation of minute-by-minute power load data within the preset historical time period, the meteorological feature data characterizes the lag of minute-by-minute power load data relative to minute-by-minute meteorological data within the preset historical time period, and the spatial feature data characterizes the impact of the geographical location of the distributed photovoltaic power station on minute-by-minute power load data within the preset historical time period; and trains the original dynamic recurrent neural network model using the time-series feature data, meteorological feature data, and spatial feature data to obtain the target prediction model. The terminal 102 may be, but is not limited to, a data storage device in each distributed power station of a distributed photovoltaic system. The server 104 may be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services.

[0047] In one exemplary embodiment, such as Figure 2 As shown, a training method for a photovoltaic output prediction model of a distributed photovoltaic system based on minute-level data is provided, and this method is applied to... Figure 1 Taking server 102 as an example, the explanation includes the following steps S202 to S208. Wherein:

[0048] S202, Obtain photovoltaic output data related to the distributed photovoltaic system; the photovoltaic output data includes minute-by-minute power load data and minute-by-minute meteorological data for the distributed photovoltaic power stations located in different geographical locations within a preset historical time period.

[0049] The preset historical time period refers to the historical data time interval pre-set for model training based on the ultra-short-term forecasting needs of distributed photovoltaic power. Its length can be flexibly configured according to the actual forecasting scenario (such as predicting output in the next 15 minutes to 4 hours), with the aim of providing the model with sufficient historical samples to learn the changing patterns of photovoltaic output. The per-minute power load data refers to power load-related data collected at 1-minute intervals, specifically including the per-minute power generation data of the distributed photovoltaic power station itself, the per-minute grid load data of the distribution network (i.e., the real-time load of photovoltaic power injected into the distribution network), and the per-minute power consumption data of the user side (i.e., the actual power consumed by users), to accurately capture the minute-level fluctuation characteristics of photovoltaic output. The per-minute meteorological data refers to meteorological parameter data closely related to photovoltaic output acquired at a 1-minute acquisition frequency, including key indicators such as irradiance, temperature, and humidity, as well as satellite remote sensing cloud cover data acquired at a resolution ≤1km² and an update frequency of 5 minutes, used to comprehensively reflect the real-time impact of meteorological factors such as sunlight and cloud cover changes on photovoltaic output.

[0050] For example, smart meters are used to collect real-time data on the power generation of distributed photovoltaic power stations, the grid load data on the distribution network side, and the electricity load data on the user side; irradiance, temperature, humidity, and other data are collected every minute using professional meteorological monitoring equipment; and high-resolution cloud cover data is obtained through satellite remote sensing technology. Finally, these data are integrated to form complete photovoltaic power output-related data required for model training.

[0051] S204, preprocess the photovoltaic output-related data to obtain the preprocessed data.

[0052] Preprocessing can refer to operations such as outlier removal, missing value imputation, normalization, and data format conversion, and is not limited here.

[0053] S206, feature extraction is performed on the preprocessed data to obtain time-series feature data, meteorological feature data and spatial feature data; the time-series feature data represents the trend and fluctuation of the power load data per minute within a preset historical time period, the meteorological feature data represents the lag of the power load data per minute relative to the meteorological data per minute within a preset historical time period, and the spatial feature data represents the impact of the geographical location of the distributed photovoltaic power station on the power load data per minute within a preset historical time period.

[0054] The time-series characteristic data refers to the feature data extracted from preprocessed 1-minute high-frequency power load data (including distributed photovoltaic power output data, distribution network load data, and user-side power load data), which can characterize the short-term trend and fluctuation of power load within a preset historical time period. Specifically, it includes the moving average and volatility. The moving average is obtained by moving a fixed-length window over the time series and calculating the average value of the data within the window, reflecting the short-term trend of load data changes. The volatility is obtained by calculating the fluctuation amplitude of load data within a certain time range, reflecting the fluctuation characteristics of load data. The meteorological characteristic data refers to the feature data extracted from preprocessed 1-minute meteorological monitoring data (including irradiance, temperature, and humidity) and minute-level satellite remote sensing cloud cover data, used to characterize the lagged correlation between power load data and meteorological data within a preset historical time period, mainly the lagged correlation between irradiance and photovoltaic power output. By analyzing this lagged correlation, the time delay relationship of the impact of irradiance changes on photovoltaic power output can be clarified, thereby quantifying the lagged effect of meteorological factors on load data. Spatial feature data refers to the characteristic data extracted from preprocessed multi-geographical distributed photovoltaic power station data, which can characterize the impact of geographical location differences on power load data within a preset historical time period. Specifically, it is reflected in the output correlation of distributed nodes. By calculating the output correlation between different distributed photovoltaic nodes (i.e., photovoltaic power stations in different geographical locations), the degree of output correlation caused by different geographical locations (affected by differences in environmental factors such as sunlight and temperature) can be reflected, thereby quantifying the impact of geographical location on load data.

[0055] For example, for electricity load data, short-term trend characteristics are obtained by calculating the moving average by moving a fixed-length window over the time series, and volatility characteristics are obtained by calculating the volatility over a certain time range, thus forming time series characteristic data; for meteorological data and electricity load data, the time delay relationship between irradiance and photovoltaic output is determined by analyzing the lag correlation between irradiance and photovoltaic output, thus forming meteorological characteristic data; for distributed photovoltaic power station data in multiple geographical locations, the impact of geographical location differences on load data is quantified by calculating the output correlation between different distributed nodes (power stations), thus forming spatial characteristic data.

[0056] S208 uses time-series feature data, meteorological feature data, and spatial feature data to train the original dynamic recurrent neural network model to obtain the target prediction model.

[0057] For example, based on the original dynamic recurrent neural network (including input layer, hidden layer, and output layer), the extracted time-series feature data, meteorological feature data, and spatial feature data are fused to form a multi-dimensional input sample, which is then input into the network input layer. The hidden layer can adopt an improved GRU (Gated Recurrent Unit) structure with an attention mechanism. This structure can dynamically identify and weight the features that have a key impact on the prediction results, effectively handling long-term dependencies and complex correlations in the data. During training, the model parameters can be optimized based on the loss function and parameter optimization algorithm. Through continuous iterative training, the model loss function value is stabilized and meets the preset accuracy requirements, so that the obtained target prediction model can output the photovoltaic power output sequence for the next 15 minutes to 4 hours, and the output resolution is 1 minute (that is, the target prediction model can predict the photovoltaic power output in each 1-minute time interval in the future, meeting the needs of ultra-short-term power prediction).

[0058] The training method for the photovoltaic output prediction model of the distributed photovoltaic system based on minute-level data uses 1-minute high-frequency power load and meteorological data as a foundation to capture minute-level fluctuations in photovoltaic output, thus solving the problem of insufficient timeliness of traditional data. By preprocessing to ensure data quality, it extracts three types of features—time series, meteorology, and space—to adapt to the strong volatility and spatial heterogeneity of photovoltaics, overcoming the defect of poor model adaptability, and significantly improving the accuracy and real-time performance of ultra-short-term predictions. This provides comprehensive information for grid dispatch, helps optimize resource allocation, improves the stability of new energy grid connection, and enhances the efficiency of green energy consumption.

[0059] In an exemplary embodiment, a target prediction model is obtained by training an original dynamic recurrent neural network model using time-series feature data, meteorological feature data, and spatial feature data, including:

[0060] By substituting time-series feature data, meteorological feature data, and spatial feature data into the original dynamic recurrent neural network model, and adjusting the parameters of the original dynamic recurrent neural network model through an adaptive momentum estimation algorithm and a learning rate decay strategy, the performance index of the original dynamic recurrent neural network model reaches the preset conditions, thereby obtaining the target prediction model.

[0061] For example, the extracted time-series feature data, meteorological feature data, and spatial feature data are fused and then fed into the original dynamic recurrent neural network model (including an input layer, an improved GRU hidden layer with an attention mechanism, and an output layer) to form the input sample for the multidimensional time series. During training, the Adaptive Momentum Estimation (AdamW) algorithm is selected as the optimization algorithm. As an improved version of the Adam algorithm, this algorithm can adaptively adjust the learning rate of each parameter of the model during the optimization process, improving training efficiency and model convergence speed. At the same time, combined with the learning rate decay strategy, a larger learning rate is used in the early stage of training to drive the model to converge quickly to the vicinity of the optimal solution, and the learning rate is gradually reduced in the later stage of training to avoid the model oscillating around the optimal solution and to ensure the model's generalization ability. During the training process, the model parameters are continuously iterated and adjusted until the model's performance indicators (such as prediction accuracy and loss function value) reach the preset conditions, thus obtaining the target prediction model.

[0062] In this embodiment, by fusing time-series, meteorological, and spatial feature data into the original dynamic recurrent neural network model with an improved GRU hidden layer containing an attention mechanism, the complex variation patterns of distributed photovoltaic power output can be accurately captured by fully utilizing multi-dimensional features, adapting to its strong volatility and spatial heterogeneity. The adaptive momentum estimation (AdamW) algorithm is selected, which can adaptively adjust the learning rate of each parameter of the model, effectively improving training efficiency and model convergence speed. Combined with the learning rate decay strategy, it ensures that the model quickly approaches the optimal solution in the early stage of training, while avoiding the model oscillating around the optimal solution in the later stage of training, significantly enhancing the model's generalization ability. By continuously iterating and adjusting the parameters until the performance indicators reach the target, the final target prediction model can more accurately output the ultra-short-term photovoltaic power output sequence, greatly improving the prediction accuracy, and providing reliable model support for efficient grid dispatch, ensuring the safe and stable operation of the grid, and improving the efficiency of green energy consumption.

[0063] In one exemplary embodiment, such as Figure 3 As shown, the performance indicators include the target loss function value. The minimum value of the target loss function corresponds to a preset condition. The target loss function is a composite function composed of the mean squared error function and the mean absolute error function. Time-series feature data, meteorological feature data, and spatial feature data are substituted into the original dynamic recurrent neural network model. The parameters of the original dynamic recurrent neural network model are adjusted through an adaptive momentum estimation algorithm and a learning rate decay strategy so that the performance indicators of the original dynamic recurrent neural network model reach the preset conditions, thereby obtaining the target prediction model, including:

[0064] S302 uses an adaptive momentum estimation algorithm and a learning rate decay strategy to calculate the minimum value of the target loss function of the original dynamic recurrent neural network model.

[0065] S304. Adjust the parameters of the original dynamic recurrent neural network model according to the minimum value of the target loss function to obtain the target prediction model.

[0066] The target loss function is a mathematical function used to measure the error between the prediction results of the original dynamic recurrent neural network model and the actual distributed photovoltaic power output data. Specifically, it is a composite function composed of the mean squared error (MSE) and the mean absolute error (MAE). The MSE, by averaging the squares of the differences between the predicted and actual values, amplifies the impact of larger errors and provides stricter control over the details of the model's prediction accuracy. The MAE, by averaging the absolute values ​​of the differences between the predicted and actual values, is less sensitive to outliers and better reflects the overall level of prediction error. The composite loss function, combining these two functions, balances the detailed accuracy of the error with overall stability, avoiding the error measurement bias caused by a single loss function. This provides a more comprehensive and accurate optimization direction for model parameter adjustment, ensuring that the final target prediction model reduces both the overall prediction error and the occurrence of extreme errors, meeting the dual requirements of accuracy and reliability for ultra-short-term distributed photovoltaic prediction.

[0067] For example, the preprocessed extracted time-series feature data, meteorological feature data, and spatial feature data are fused to form a multi-dimensional time-series input sample, which is then substituted into the original dynamic recurrent neural network model. The model training process is initiated, employing the Adaptive Momentum Estimation Algorithm (AdamW) as the parameter optimization algorithm: based on the gradient descent principle, the learning rate is adaptively adjusted according to the gradient changes of each model parameter. For example, a larger learning rate is set for parameters with small gradient changes (parameters with less impact on prediction results) to accelerate updates, while a smaller learning rate is set for parameters with large gradient changes (parameters with greater impact on prediction results) to avoid excessive update amplitude, thereby improving training efficiency and model convergence speed. Simultaneously, a learning rate decay strategy is employed: a larger initial learning rate is set in the early stages of training to drive the model to converge quickly to near a better solution (reducing early training time and avoiding getting trapped in local optima). As the number of training iterations increases, the learning rate is gradually reduced to prevent the model from oscillating around the optimal solution. During training, the error between the model's predicted values ​​and the actual photovoltaic power output data is continuously calculated using a composite loss function: the mean squared error function amplifies the weight of larger errors, ensuring the model focuses on optimizing predictions with significant deviations; the mean absolute error function balances the overall error level, avoiding over-sensitivity to outliers due to reliance solely on the mean squared error. The AdamW algorithm continuously adjusts model parameters (such as the gating parameters of the hidden layer GRU and the weight parameters of the attention mechanism) based on the gradient feedback of the loss function, and dynamically adjusts the optimization step size using a learning rate decay strategy, iterating this "calculate loss - adjust parameters" process repeatedly. When the minimum value of the composite loss function is calculated using the above optimization method (i.e., the model prediction error reaches the preset minimum threshold, meeting performance requirements), parameter iteration stops. At this point, the target prediction model is determined based on the model parameter state corresponding to this minimum value.

[0068] In this embodiment, a composite loss function consisting of MSE and MAE is adopted to balance the details of prediction accuracy with the overall stability of error, avoiding the bias of a single loss function. The AdamW algorithm is used to adaptively adjust the parameter learning rate to improve training efficiency and convergence speed, and a learning rate decay strategy is used to prevent model oscillation and enhance generalization ability. The three types of feature data are fused and iteratively optimized to minimize the loss function. The resulting target prediction model can accurately output minute-level ultra-short-term photovoltaic power output sequences, providing reliable support for grid dispatch, safe operation and green energy consumption.

[0069] In one exemplary embodiment, such as Figure 4 As shown, the photovoltaic output data of the distributed photovoltaic system is preprocessed to obtain the preprocessed data, including:

[0070] S402, the sliding window method is used to remove outliers in the photovoltaic power output data.

[0071] S404. Normalize the remaining photovoltaic output data after removing outliers to obtain preprocessed data.

[0072] For example, a sliding window method is used to remove outliers from the collected photovoltaic output-related data (including 1-minute distributed photovoltaic output data, distribution network load data, user-side electricity load data, 1-minute meteorological monitoring data, and processed minute-level satellite remote sensing cloud cover data). During operation, the sliding window moves across the data sequence at preset steps. For each dataset within the window, a reasonable anomaly threshold is first set based on the normal distribution range of the data or business experience (e.g., data deviating from the average value of the data within the window is considered anomaly). Then, by comparing the relationship between each data point within the window and the threshold, outlier data exceeding the threshold range is filtered out and removed. This effectively removes outliers caused by equipment failure, data transmission interference, etc., ensuring the accuracy and reliability of the data and laying a high-quality data foundation for subsequent processing. After outlier removal, the remaining photovoltaic output-related data is normalized using the Z-score standardization method. According to the Z-score standardization formula: z=(x-μ) / σ, (where x is the original data after removing outliers, μ is the overall mean of the data type, and σ is the overall standard deviation of the data type), different types of data are processed separately: for example, for 1-minute photovoltaic power output data, the mean is first calculated. with standard deviation Then, each data point is substituted into the formula to obtain the standardized output data. For meteorological data such as irradiance and temperature, their respective means and standard deviations are calculated and standardized. Through this processing, data with significantly different dimensions (such as photovoltaic output in "kilowatts", temperature in "degrees Celsius", and irradiance in "watts per square meter") are converted into standardized data with unified dimensions. This completely eliminates the impact of dimensional differences on the accuracy of subsequent feature extraction and the convergence speed of model training, ensuring that different types of data can be effectively utilized within the same analytical framework, ultimately obtaining preprocessed data that meets the requirements of subsequent processes.

[0073] In this embodiment, outliers are eliminated using the sliding window method, which can identify and remove abnormal data caused by equipment failure, data transmission interference, etc., effectively ensuring the accuracy and reliability of the data. This lays a high-quality data foundation for subsequent feature extraction and model training, and avoids interference from abnormal data in subsequent processes. Through normalization processing, various types of data are converted into standardized data with a unified dimension, eliminating the impact of dimensional differences on the accuracy of feature extraction and the convergence speed of model training. This ensures that different types of data can be effectively utilized under the same analytical framework, thereby improving the efficiency and prediction accuracy of subsequent model training and providing strong support for the construction of distributed photovoltaic ultra-short-term prediction models.

[0074] In one exemplary embodiment, such as Figure 5 As shown, before training the original dynamic recurrent neural network model using time-series feature data, meteorological feature data, and spatial feature data, the method further includes:

[0075] S502 clusters time-series characteristic data, meteorological characteristic data, and spatial characteristic data belonging to the same time window to obtain multiple clusters.

[0076] S504 uses multiple clusters as training sample sets and substitutes them into the original dynamic recurrent neural network model.

[0077] For example, using a 30-minute time window, from the preprocessed time-series, meteorological, and spatial feature data, we select the three types of feature data belonging to the same 30-minute time window. This ensures that each group of data to be clustered corresponds to complete feature information within the same time period, fully reflecting the operating status and environmental influencing factors of distributed photovoltaic systems within that time period. Clustering is then performed on the time-series, meteorological, and spatial feature data belonging to the same 30-minute time window: Clustering algorithms (such as K-means, DBSCAN, etc.) are used to group the feature data groups corresponding to multiple 30-minute time windows. Based on the similarity of the feature data (such as similar load fluctuation trends, similar meteorological influence patterns, and similar node output correlation), they are divided into multiple clusters. This ensures that the feature data groups within each cluster have highly similar distribution characteristics and variation patterns, effectively reducing data redundancy and highlighting the differences in data features under different operating scenarios. After obtaining multiple clusters through clustering, these clusters are directly used as training sample sets and substituted into the original dynamic recurrent neural network model. Since each cluster originates from the clustering of multidimensional feature data within the same 30-minute time window, the model can learn the changing patterns and interrelationships of the data within the 30-minute time period. Furthermore, clustering groups samples with similar features into one category, making the structure of the training sample set more regular and the features more focused. This means that the model can learn separately for the operating scenarios corresponding to different clusters during training, reducing cross-interference between data from different scenarios and more accurately capturing the correlation patterns of photovoltaic power output under the same scenario. This lays the foundation for subsequent efficient and high-precision model training based on this training sample set.

[0078] In this embodiment, the three types of feature data—time series, meteorological, and spatial—are clustered within the same window. This ensures that the model learns the data change patterns and correlations within the same time window, and also reduces data redundancy and highlights scene differences by clustering based on feature similarity. Substituting these clusters into the model as training sample sets makes the sample structure regular and the features focused, reducing cross-interference between scene data and helping the model accurately capture the correlation patterns of photovoltaic power output under similar scenarios. This lays the foundation for efficient training and improved accuracy of ultra-short-term predictions.

[0079] In one exemplary embodiment, such as Figure 6 As shown, this application provides a method for predicting the photovoltaic output of a distributed photovoltaic system, the method comprising:

[0080] S602, Obtain photovoltaic output-related data of the distributed photovoltaic system; the photovoltaic output-related data includes minute-by-minute power load data and minute-by-minute meteorological data for distributed photovoltaic power stations located in different geographical locations within a preset historical time period; the preset historical time period is a period before the time period to be predicted;

[0081] S604, preprocesses the photovoltaic output-related data to obtain preprocessed data;

[0082] S606, feature extraction is performed on the preprocessed data to obtain time-series feature data, meteorological feature data and spatial feature data; the time-series feature data represents the trend and fluctuation of the power load data per minute within a preset historical time period, the meteorological feature data represents the lag of the power load data per minute relative to the meteorological data per minute within a preset historical time period, and the spatial feature data represents the impact of the geographical location of the distributed photovoltaic power station on the power load data per minute within a preset historical time period.

[0083] S608, input the time series feature data, meteorological feature data and spatial feature data into the target prediction model to obtain the photovoltaic power output within the time period to be predicted; the target prediction model is trained using the training method of the photovoltaic power output prediction model of the distributed photovoltaic system based on minute-level data in the above embodiment.

[0084] The specific implementation process and beneficial effects of the photovoltaic output prediction method for the above-mentioned distributed photovoltaic system are similar to the model training process described above, and will not be repeated here.

[0085] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0086] Based on the same inventive concept, this application also provides a training device for a photovoltaic power output prediction model of a distributed photovoltaic system based on minute-level data, used to implement the training method for the photovoltaic power output prediction model of a distributed photovoltaic system based on minute-level data described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more training device embodiments for a photovoltaic power output prediction model of a distributed photovoltaic system based on minute-level data provided below can be found in the limitations of the training method for the photovoltaic power output prediction model of a distributed photovoltaic system based on minute-level data described above, and will not be repeated here.

[0087] In one exemplary embodiment, such as Figure 7 As shown, a training device for a photovoltaic output prediction model of a distributed photovoltaic system based on minute-level data is provided, including: a data acquisition module 702, a preprocessing module 704, a feature extraction module 706, and a training module 708, wherein:

[0088] The data acquisition module 702 is used to acquire photovoltaic output-related data of the distributed photovoltaic system; the photovoltaic output-related data includes minute-by-minute power load data and minute-by-minute meteorological data for a preset historical time period corresponding to distributed photovoltaic power stations in different geographical locations;

[0089] Preprocessing module 704 is used to preprocess the photovoltaic output-related data to obtain preprocessed data;

[0090] The feature extraction module 706 is used to extract features from the preprocessed data to obtain time-series feature data, meteorological feature data, and spatial feature data. The time-series feature data represents the trend and fluctuation of the minute-by-minute power load data within the preset historical time period. The meteorological feature data represents the lag of the minute-by-minute power load data relative to the minute-by-minute meteorological data within the preset historical time period. The spatial feature data represents the impact of the geographical location of the distributed photovoltaic power station on the minute-by-minute power load data within the preset historical time period.

[0091] The training module 708 is used to train the original dynamic recurrent neural network model using the time-series feature data, the meteorological feature data, and the spatial feature data to obtain the target prediction model.

[0092] In an exemplary embodiment, the training module 708 described above includes:

[0093] The training unit is used to input the time-series feature data, meteorological feature data, and spatial feature data into the original dynamic recurrent neural network model, and adjust the parameters of the original dynamic recurrent neural network model through an adaptive momentum estimation algorithm and a learning rate decay strategy, so that the performance index of the original dynamic recurrent neural network model reaches the preset conditions, thereby obtaining the target prediction model.

[0094] In an exemplary embodiment, the performance metric includes a target loss function value, the minimum value of which corresponds to the preset condition, and the target loss function is a composite function consisting of a mean squared error function and a mean absolute error function; the training unit includes:

[0095] The loss function value determination unit is used to calculate the minimum value of the target loss function of the original dynamic recurrent neural network model using an adaptive momentum estimation algorithm and a learning rate decay strategy.

[0096] The parameter adjustment unit is used to adjust the parameters of the original dynamic recurrent neural network model according to the minimum value of the target loss function, thereby obtaining the target prediction model.

[0097] In one exemplary embodiment, the preprocessing module 704 includes:

[0098] The elimination unit is used to eliminate outliers in the photovoltaic output-related data using a sliding window method.

[0099] The normalization unit is used to normalize the photovoltaic output-related data remaining after removing outliers to obtain the preprocessed data.

[0100] In an exemplary embodiment, the training module 708 further includes:

[0101] The classification unit is used to cluster the time-series feature data, meteorological feature data, and spatial feature data that belong to the same time window to obtain multiple clusters.

[0102] The input unit is used to substitute multiple clusters as training sample sets into the original dynamic recurrent neural network model.

[0103] The modules in the training device for the photovoltaic output prediction model of the distributed photovoltaic system based on minute-level data can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0104] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 8 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores photovoltaic output-related data. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a training method for a photovoltaic output prediction model of a distributed photovoltaic system based on minute-level data.

[0105] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0106] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0107] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0108] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0109] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0110] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0111] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for training a photovoltaic (PV) power output prediction model of a distributed PV system based on minute-level data, characterized in that, The method includes: Acquire photovoltaic output data related to the distributed photovoltaic system; the photovoltaic output data includes minute-by-minute power load data and minute-by-minute meteorological data for distributed photovoltaic power stations located in different geographical locations within a preset historical time period; The photovoltaic output-related data are preprocessed to obtain preprocessed data; Feature extraction is performed on the preprocessed data to obtain time-series feature data, meteorological feature data, and spatial feature data. The time-series feature data represents the trend and fluctuation of the minute-by-minute power load data within the preset historical time period. The meteorological feature data represents the lag of the minute-by-minute power load data relative to the minute-by-minute meteorological data within the preset historical time period. The spatial feature data represents the impact of the geographical location of the distributed photovoltaic power station on the minute-by-minute power load data within the preset historical time period. The original dynamic recurrent neural network model is trained using the time-series feature data, the meteorological feature data, and the spatial feature data to obtain the target prediction model. 2.The method of claim 1, wherein, The step of training the original dynamic recurrent neural network model using the time-series feature data, the meteorological feature data, and the spatial feature data to obtain the target prediction model includes: The time-series feature data, meteorological feature data, and spatial feature data are substituted into the original dynamic recurrent neural network model. The parameters of the original dynamic recurrent neural network model are adjusted by an adaptive momentum estimation algorithm and a learning rate decay strategy so that the performance index of the original dynamic recurrent neural network model reaches the preset conditions, thereby obtaining the target prediction model. 3.The method of claim 2, wherein, The performance index includes a target loss function value, the minimum value of which corresponds to the preset condition. The target loss function is a composite function composed of a mean squared error function and a mean absolute error function. The process involves substituting the time-series feature data, meteorological feature data, and spatial feature data into the original dynamic recurrent neural network model, adjusting the parameters of the original dynamic recurrent neural network model using an adaptive momentum estimation algorithm and a learning rate decay strategy, so that the performance index of the original dynamic recurrent neural network model reaches the preset condition, thereby obtaining the target prediction model, including: The minimum value of the target loss function of the original dynamic recurrent neural network model is calculated using an adaptive momentum estimation algorithm and a learning rate decay strategy. The parameters of the original dynamic recurrent neural network model are adjusted according to the minimum value of the target loss function to obtain the target prediction model. 4.The method of claim 1, wherein, The preprocessing of the photovoltaic output-related data of the distributed photovoltaic system to obtain preprocessed data includes: The sliding window method was used to remove outliers from the photovoltaic power output data. The photovoltaic output data remaining after removing outliers is normalized to obtain the preprocessed data. 5.The method of claim 1, wherein, Before the step of training the original dynamic recurrent neural network model using the time-series feature data, the meteorological feature data, and the spatial feature data, the method further includes: The time-series feature data, meteorological feature data, and spatial feature data belonging to the same time window are clustered to obtain multiple clusters; Multiple clusters are used as training sample sets and substituted into the original dynamic recurrent neural network model.

6. A photovoltaic power output prediction method for a distributed photovoltaic system, characterized by, The method includes: Acquire photovoltaic output-related data of distributed photovoltaic systems, including minute-by-minute power load data and minute-by-minute meteorological data for distributed photovoltaic power stations located in different geographical locations within a preset historical time period; the preset historical time period is a period before the time period to be predicted; The photovoltaic output-related data are preprocessed to obtain preprocessed data; Feature extraction is performed on the preprocessed data to obtain time-series feature data, meteorological feature data, and spatial feature data. The time-series feature data represents the trend and fluctuation of the minute-by-minute power load data within the preset historical time period. The meteorological feature data represents the lag of the minute-by-minute power load data relative to the minute-by-minute meteorological data within the preset historical time period. The spatial feature data represents the impact of the geographical location of the distributed photovoltaic power station on the minute-by-minute power load data within the preset historical time period. The time-series feature data, meteorological feature data, and spatial feature data are input into the target prediction model to obtain the photovoltaic output within the time period to be predicted; the target prediction model is trained using the training method of the photovoltaic output prediction model of a distributed photovoltaic system based on minute-level data as described in any one of claims 1 to 5. 7.A device for training a photovoltaic (PV) power output prediction model of a distributed PV system based on minute-level data, characterized in that, The device includes: The data acquisition module is used to acquire photovoltaic output-related data of the distributed photovoltaic system; the photovoltaic output-related data includes minute-by-minute power load data and minute-by-minute meteorological data for distributed photovoltaic power stations located in different geographical locations within a preset historical time period; The preprocessing module is used to preprocess the photovoltaic output-related data to obtain preprocessed data; The feature extraction module is used to extract features from the preprocessed data to obtain time-series feature data, meteorological feature data, and spatial feature data. The time-series feature data represents the trend and fluctuation of the minute-by-minute power load data within the preset historical time period. The meteorological feature data represents the lag of the minute-by-minute power load data relative to the minute-by-minute meteorological data within the preset historical time period. The spatial feature data represents the impact of the geographical location of the distributed photovoltaic power station on the minute-by-minute power load data within the preset historical time period. The training module is used to train the original dynamic recurrent neural network model using the time-series feature data, the meteorological feature data, and the spatial feature data to obtain the target prediction model.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

10. A computer program product comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images