A Distributed Photovoltaic Power Prediction Method and System Based on Adaptive Learning

Abstract

This invention discloses a distributed photovoltaic (PV) power prediction method and system based on adaptive learning, relating to the field of distributed PV power prediction. The method includes: using the current time node as the zero point, collecting historical power data, component temperature data, and inverter efficiency data within a preset time period to establish a backtracking period dataset; sending the cloud map prediction data and the backtracking period dataset to a mutation-aware scheduler; after matching and activating the predictors of a multi-granularity prediction model cluster according to the prediction granularity, extracting a mutation point heatmap based on the cloud map prediction data and the backtracking period dataset; performing mutation attention alignment on the predictors, and then performing PV power prediction; and outputting the PV power prediction result after self-verification and correction of the dual-granularity prediction results. This invention addresses the technical problem that existing power prediction methods cannot meet the grid's dual requirements for prediction accuracy and flexibility, achieving the technical effect of improving the accuracy and flexibility of power prediction.

Description

Technical Field

[0001] This application relates to the field of distributed photovoltaic power prediction, and in particular to a distributed photovoltaic power prediction method and system based on adaptive learning. Background Technology

[0002] Accurate forecasting of distributed photovoltaic (PV) power is a core component supporting flexible grid dispatch and ensuring high-proportion renewable energy consumption. Its forecasting accuracy directly impacts the safe, stable operation and economic efficiency of the power system. However, PV power is significantly affected by sudden weather changes (such as rapid cloud movement and localized showers), and the forecasting requirements vary greatly across different time scales (e.g., short-term minute-level fluctuations versus long-term day-level trends). Currently, the main approach to addressing this issue is to use a single forecasting model with fixed time granularity or simply superimpose short- and long-period models. These methods lack real-time sensing capabilities for power fluctuations and fail to establish a dynamic collaborative correction mechanism for short- and long-period forecasting results. Consequently, when weather changes rapidly, short-period forecasts are easily affected by noise and deviate from the true trend, while long-period forecasts lose real-time performance due to neglecting local details. Furthermore, the two methods cannot complement each other when operating independently.

[0003] Currently, distributed photovoltaic power prediction technologies face the challenge of failing to meet the grid's dual requirements for prediction accuracy and flexibility. Summary of the Invention

[0004] This application provides a distributed photovoltaic power prediction method and system based on adaptive learning. It employs a method that collects historical power, module temperature, and inverter efficiency data based on the current time point to construct a backtracking periodic dataset. Simultaneously, it reads cloud map prediction data. These two types of data are input into a mutation-aware scheduler to generate first and second prediction granularities. A matching predictor is activated based on the prediction granularity, and a mutation point heatmap is extracted and attention aligned. Photovoltaic power prediction is then performed to obtain dual-granularity results. Finally, the final prediction result is output through self-verification correction. These techniques address the technical problem of existing distributed photovoltaic power prediction methods failing to meet the grid's dual requirements for prediction accuracy and flexibility, thus achieving the technical effect of improving the accuracy and flexibility of distributed photovoltaic power prediction.

[0005] This application provides a distributed photovoltaic power prediction method based on adaptive learning, comprising: taking the current time node as the zero point, performing data collection of historical power data, component temperature data, and inverter efficiency data within a preset time period to establish a backtracking period dataset; after reading cloud map prediction data, sending the cloud map prediction data and the backtracking period dataset to a mutation-aware scheduler, and outputting prediction granularity, wherein the prediction granularity includes a first prediction granularity and a second prediction granularity, the second prediction granularity being constructed based on the first prediction granularity and being larger than the first prediction granularity; after performing predictor matching activation of a multi-granularity prediction model cluster according to the prediction granularity, extracting a mutation point heatmap according to the cloud map prediction data and the backtracking period dataset; after using the mutation point heatmap to perform mutation attention alignment of the activated predictors, performing photovoltaic power prediction to establish a dual-granularity prediction result; and after performing self-verification correction on the dual-granularity prediction result, outputting the photovoltaic power prediction result.

[0006] In a possible implementation, the cloud image prediction data and the backtracking period dataset are sent to a mutation-aware scheduler, which outputs the prediction granularity and performs the following processing: After the mutation-aware scheduler receives the cloud image prediction data and the backtracking period data, it performs mutation feature perception as follows: After sorting the cloud image prediction data by time, spatial features are extracted using a convolutional image processing algorithm. The spatial features include cloud edge density gradient map, cloud edge displacement velocity vector, and temporal change rate of the occlusion coverage area. A local occlusion mutation factor map is established using the spatial features. The backtracking period dataset is subjected to time-dimensional mutation analysis based on a sliding window to extract a perturbation feature sequence. After standardization processing based on the perturbation feature sequence and the local occlusion mutation factor map, feature splicing and tensor fusion are performed to establish a mutation exponential tensor. The prediction granularity is output using the mutation exponential tensor.

[0007] In a possible implementation, the backtracking period dataset is subjected to time-dimensional mutation analysis based on a sliding window to extract disturbance feature sequences, and the following processing is performed: Direct volatility feature extraction is performed within the sliding window to establish a multi-dimensional trend disturbance factor group, which includes a local volatility factor, a trend slope mutation factor, a range disturbance ratio factor, and a frequency domain disturbance energy factor; Induced disturbance feature extraction is performed within the sliding window to establish an induced disturbance factor group, which includes a thermal mutation potential factor, a performance drift index factor, and a probability disturbance memory factor; Disturbance feature sequences are extracted based on the multi-dimensional trend disturbance factor group and the induced disturbance factor group.

[0008] In a possible implementation, the mutation exponent tensor is used to output the prediction granularity, and the following processing is performed: the mutation exponent tensor is used as a tensor input, and the prediction probabilities of multiple candidate granularities are output based on a lightweight particle prediction network; after filtering the candidate granularities according to the prediction probabilities, the first prediction granularity is output; the first prediction granularity is optimized by upward abstraction of the global verification scale configuration, and the second prediction granularity is established using the configuration optimization result.

[0009] In a possible implementation, after self-verification correction of the dual-granularity prediction results, the photovoltaic power prediction results are output, and the following processing is performed: interpolation and alignment of the dual-granularity prediction results are performed to construct a prediction value difference tensor on the same time axis; the consistency coefficient and offset drift index are calculated using the prediction value difference tensor to generate a granularity result consistency score; if the granularity result consistency score meets the activation condition, the residual correction network is activated; the first prediction granularity result is corrected using the residual correction network according to the prediction value difference tensor and the enhancement factor, and then output as the photovoltaic power prediction result.

[0010] In a possible implementation, the mutation attention alignment of the activated predictor is performed using the mutation point heatmap, and the following processing is performed: threshold segmentation is performed on the mutation point heatmap to extract a set of mutation key points; mutation value attention weighting is applied to the mutation key point mapping of each mutation key point using the set of mutation key points to complete the predictor mutation attention initialization at the first prediction granularity; and gradual attention weighting is applied to each mutation point in the set of mutation key points to complete the predictor mutation attention initialization at the second prediction granularity.

[0011] In a possible implementation, after outputting the photovoltaic power prediction result, the following processing is performed: establishing a verification time node based on the first prediction granularity; reading photovoltaic power at the verification time node to establish verification data; using the verification data and the photovoltaic power prediction result to perform deviation verification and establish deviation feedback; and using the deviation feedback as correction data to perform adaptive correction compensation for the prediction.

[0012] In a possible implementation, the deviation feedback is used as correction data to perform adaptive correction compensation for prediction, and the following processing is also performed: continuous feedback monitoring of the deviation feedback is performed, and it is determined whether the continuous feedback monitoring result meets the trigger threshold; if the continuous feedback monitoring result meets the trigger threshold, an online incremental correction mechanism is triggered to perform prediction training reconstruction.

[0013] In a possible implementation, the cloud map prediction data and the backtracking periodic dataset are sent to the mutation-aware scheduler, the prediction granularity is output, and the following processing is performed: the prediction granularity is judged by a granularity reliability index; if the index judgment result is a failure result, the redundant prediction compensation mechanism is activated, the prediction granularity is reconstructed, and redundant prediction is performed.

[0014] This application also provides a distributed photovoltaic power prediction system based on adaptive learning, comprising: a historical data acquisition module, used to collect historical power data, component temperature data, and inverter efficiency data within a preset time period, using the current time node as the zero point, and establish a backtracking period dataset; a granular prediction module, used to read cloud map prediction data, send the cloud map prediction data and the backtracking period dataset to a mutation-aware scheduler, and output prediction granularity, wherein the prediction granularity includes a first prediction granularity and a second prediction granularity, the second prediction granularity being constructed based on the first prediction granularity and being larger than the first prediction granularity; a mutation point heatmap extraction module, used to extract a mutation point heatmap based on the cloud map prediction data and the backtracking period dataset after matching and activating the predictors of a multi-granularity prediction model cluster according to the prediction granularity; a photovoltaic power prediction module, used to perform photovoltaic power prediction after aligning the mutation attention of the activated predictors using the mutation point heatmap, and establish a dual-granularity prediction result; and a self-verification correction module, used to perform self-verification correction on the dual-granularity prediction result and output a photovoltaic power prediction result.

[0015] This application proposes a distributed photovoltaic (PV) power prediction method and system based on adaptive learning. First, using the current time node as the zero point, historical power data, component temperature data, and inverter efficiency data within a preset time period are collected to establish a backtracking period dataset. Next, after reading cloud map prediction data, the cloud map prediction data and the backtracking period dataset are sent to a mutation-aware scheduler, which outputs the prediction granularity. This prediction granularity includes a first prediction granularity and a second prediction granularity, with the second prediction granularity built upon the first and larger than the first. Then, predictor matching and activation of a multi-granularity prediction model cluster are performed based on the prediction granularity. A mutation point heatmap is extracted from the cloud map prediction data and the backtracking period dataset. The activated predictors are then aligned with mutation attention using the mutation point heatmap before PV power prediction is performed, establishing a dual-granularity prediction result. Finally, the dual-granularity prediction result is self-verified and corrected before outputting the PV power prediction result. This achieves the technical effect of improving the accuracy and flexibility of distributed PV power prediction. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 This is a flowchart illustrating a distributed photovoltaic power prediction method based on adaptive learning, provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of the structure of a distributed photovoltaic power prediction system based on adaptive learning, provided in an embodiment of this application.

[0019] Figure labeling: Historical data acquisition module 10, granularity prediction module 20, mutation point heat map extraction module 30, photovoltaic power prediction module 40, self-verification correction module 50. Detailed Implementation

[0020] The above description is merely an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below.

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] In the following description, references to "some embodiments" describe a subset of all possible embodiments. However, it is understood that "some embodiments" can be the same or different subsets of all possible embodiments and can be combined with each other without conflict. The terms "first" and "second" are used merely to distinguish similar objects and do not represent a specific ordering of objects. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to these processes, methods, products, or devices. Unless otherwise defined, 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 application belongs. The terminology used herein is for the purpose of describing embodiments of this application only.

[0023] This application provides a distributed photovoltaic power prediction method based on adaptive learning, such as... Figure 1 As shown, the method includes:

[0024] Step S100: Using the current time node as the zero point, perform data collection of historical power data, component temperature data, and inverter efficiency data within a preset time period to establish a backtracking period dataset.

[0025] Specifically, data acquisition technology is employed to acquire historical power data (reflecting the power generation of photovoltaic modules), module temperature data (associated with power efficiency), and inverter efficiency data (the efficiency of the inverter in converting DC to AC power) within a preset time period (e.g., the past 24 hours or week) from various sensors deployed in the photovoltaic power plant, such as power sensors and temperature sensors. Data acquisition cards are used to sample and quantize the analog signals output from these sensors, converting them into digital signals. These digital signals are then transmitted to a data storage server via communication interfaces (e.g., RS-485, Ethernet) to construct a backtracking period dataset, providing historical reference data for analysis.

[0026] For example, with a preset time period of one day, in a small photovoltaic power station, a power sensor collects power values ​​every 15 minutes, a temperature sensor collects module temperature data every 30 minutes, and the inverter's built-in monitoring module records efficiency data every hour. The data acquisition system packages and stores the collected data in a server database according to the set time intervals, forming a backtracking periodic dataset containing fields such as timestamp, power value, module temperature, and inverter efficiency.

[0027] Step S200: After reading the cloud map prediction data, the cloud map prediction data and the backtracking periodic dataset are sent to the mutation-aware scheduler, and the prediction granularity is output. The prediction granularity includes a first prediction granularity and a second prediction granularity. The second prediction granularity is constructed based on the first prediction granularity, and the second prediction granularity is greater than the first prediction granularity.

[0028] Specifically, cloud imagery forecast data comes from meteorological satellites or professional meteorological service providers, acquired via network interfaces in specific data formats such as NetCDF and HDF. This data includes forecast information on cloud cover, cloud thickness, and cloud movement speed that changes over time. After acquiring the cloud imagery forecast data, it is sent along with a backtracking periodic dataset to a mutation-aware scheduler via a data bus or message queue mechanism. The mutation-aware scheduler is an intelligent scheduling module that can determine potential mutations based on the characteristics of the input data and accordingly determine the prediction granularity. The mutation-aware scheduler uses rule-based or machine learning algorithms to analyze the input data and determine the prediction granularity, dividing it into a first prediction granularity (a finer time interval, such as 15 minutes) and a second prediction granularity (a coarser time interval, such as 1 hour, formed by merging the first prediction granularity). The second prediction granularity is larger than the first.

[0029] For example, suppose cloud map prediction data forecasts cloud cover changes over the next 24 hours. The mutation-aware scheduler has built-in rules that determine a potential mutation when the cloud cover change rate exceeds a certain threshold (e.g., 10% per hour). If the cloud map prediction indicates a drastic change in cloud cover during a certain period, the first prediction granularity is set to 15 minutes, and the second prediction granularity is set to 1 hour. Conversely, if the cloud cover is predicted to change steadily, the first prediction granularity is set to 30 minutes, and the second prediction granularity is set to 2 hours, thus outputting the prediction granularity for different scenarios.

[0030] In one possible implementation, the cloud image prediction data and the backtracking period dataset are sent to a mutation-aware scheduler, which outputs the prediction granularity. Step S200 further includes, after the mutation-aware scheduler receives the cloud image prediction data and the backtracking period data, performing mutation feature perception, including step S210, whereby the cloud image prediction data is sorted temporally, and spatial features are extracted using a convolutional image processing algorithm. These spatial features include cloud edge density gradient maps, cloud edge displacement velocity vectors, and the temporal rate of change of the occlusion coverage area. Specifically, the cloud image prediction data contains cloud image images at multiple time points. These images need to be sorted temporally to ensure the correct temporal order of each image. Then, using a convolutional algorithm in image processing, the convolutional kernel slides across the image to calculate the convolutional value of each pixel and its neighborhood, thereby extracting the cloud edge density gradient map, cloud edge displacement velocity vector, and the temporal rate of change of the occlusion coverage area.

[0031] For example, cloud image prediction data includes hourly cloud images for the next 24 hours, each with a resolution of 512×512 pixels. These images are arranged chronologically and then convolved pixel-by-pixel using a 3×3 kernel. For instance, to extract the density gradient map of cloud edges, the convolution kernel can be designed as a Sobel operator that emphasizes edge features, calculating the gradient of each pixel in the horizontal and vertical directions to obtain the density changes at the cloud edges.

[0032] Step S220: A local shading abrupt change factor map is established using the spatial features. Specifically, based on extracted spatial features, such as cloud edge density gradient maps, cloud edge displacement velocity vectors, and the time change rate of the shading coverage area, a three-dimensional map is constructed. The horizontal and vertical axes of this map represent spatial locations, and the third dimension represents the intensity of the abrupt change factor. The intensity of the abrupt change factor is used to assess the extent to which changes in the cloud map may have abrupt effects on photovoltaic power by comprehensively considering multiple spatial features.

[0033] For example, based on the spatial features extracted in step S210, a 512×512×3 three-dimensional array is constructed. The first two dimensions correspond to the spatial location of the image, while the third dimension stores the cloud edge density gradient, cloud edge displacement velocity, and the temporal rate of change of the occlusion coverage area. For instance, at a location (x, y), if the cloud edge density gradient is 0.8, the cloud edge displacement velocity is 5 m / s, and the temporal rate of change of the occlusion coverage area is 0.3 / h, then the intensity of the abrupt change factor at that location can be calculated by weighting these indicators.

[0034] Step S230: Perform time-dimensional mutation analysis on the backtracking periodic dataset based on a sliding window to extract the perturbation feature sequence. Specifically, a sliding window technique is used to perform time-dimensional mutation analysis on the backtracking periodic dataset. The size of the sliding window can be determined based on the data sampling frequency and the time scale of the mutation features. Within each window, statistical characteristics of the data, such as mean, variance, and slope, are calculated to extract the perturbation feature sequence.

[0035] For example, the backtracking periodic dataset includes power data collected every 15 minutes over the past 24 hours, totaling 96 data points. A sliding window of size 5 is selected, with a window step size of 1. Within each window, the mean and variance of the power data are calculated to obtain a perturbation feature sequence. For example, the first window might have a mean of 25.5 kW and a variance of 2.3 kW. 2 The mean of the second window is 26.3 kW, and the variance is 3.1 kW. 2 And so on.

[0036] Step S240: After standardizing the perturbation feature sequence and the locally masked mutation factor map, feature splicing and tensor fusion are performed to establish a mutation exponential tensor. Specifically, the perturbation feature sequence and the locally masked mutation factor map are standardized to ensure consistent numerical ranges. Then, these features are spliced ​​to form a high-dimensional feature vector. Tensor fusion technology is used to combine these feature vectors into a mutation exponential tensor.

[0037] For example, the length of the perturbation feature sequence is 96, and each feature includes two indices: mean and variance; the size of the local occlusion mutation factor map is 512×512×4 (where 4 represents the cloud edge density gradient, cloud edge displacement velocity, time change rate of the occlusion coverage area, and mutation factor intensity). After standardizing these two features, they are concatenated into a 96×2+512×512×4 dimensional feature vector. Then, using tensor fusion technology, this feature vector is transformed into a 512×512×(4+2) mutation exponent tensor, where 4 comes from the local occlusion mutation factor map and 2 comes from the perturbation feature sequence.

[0038] Step S250: Output the prediction granularity using the mutation index tensor. Specifically, determine the prediction granularity based on the value of the mutation index tensor. Clustering algorithms or thresholding methods can be used to divide the mutation index tensor into different regions, each corresponding to a prediction granularity. For example, regions with higher mutation indices require finer prediction granularity, while regions with lower mutation indices can use coarser prediction granularity.

[0039] For example, based on the value of the mutation index tensor, the K-means clustering algorithm is used to divide it into two cluster centers, corresponding to the first prediction granularity (15 minutes) and the second prediction granularity (1 hour), respectively. For instance, regions with a mutation index above 0.5 use a 15-minute prediction granularity, while regions with a mutation index below 0.5 use a 1-hour prediction granularity.

[0040] This approach, by extracting the spatial features of cloud map prediction data and the temporal features of backtracking periodic data, can more accurately capture sudden changes in photovoltaic power, thereby improving the accuracy of prediction results.

[0041] In one possible implementation, the backtracking period dataset undergoes time-dimensional mutation analysis based on a sliding window to extract disturbance feature sequences. Step S230 further includes step S231, where direct fluctuation feature extraction is performed within the sliding window to establish a multidimensional trend disturbance factor group. This multidimensional trend disturbance factor group includes a local volatility factor, a trend slope mutation factor, a range disturbance ratio factor, and a frequency domain disturbance energy factor. Specifically, the standard deviations of power data, component temperature data, and inverter efficiency data are calculated within the sliding window to obtain the local volatility factor. Linear regression models are fitted to the power data, component temperature data, and inverter efficiency data within the sliding window to obtain regression coefficients (i.e., slopes), thus obtaining the trend slope mutation factor. The difference between the maximum and minimum values ​​of the power data, component temperature data, and inverter efficiency data is calculated within the sliding window and divided by the average value to obtain the range disturbance ratio factor. A Fast Fourier Transform (FFT) is performed on the power data, component temperature data, and inverter efficiency data within the sliding window to calculate the energy in the frequency domain, thus obtaining the frequency domain disturbance energy factor.

[0042] Step S232: Extract induced disturbance features within the sliding window and establish an induced disturbance factor group, which includes a thermal mutation potential factor, a performance drift index factor, and a probability disturbance memory factor. Specifically, calculate the component temperature change rate within the sliding window to obtain the thermal mutation potential factor. Calculate the drift degree of power data and inverter efficiency data within the sliding window, using the R-squared value of linear regression (R²). 2 The performance drift index factor is obtained by calculating the autocorrelation coefficient (ACC) of power data, component temperature data, and inverter efficiency data within the sliding window. The probability perturbation memory factor is then obtained.

[0043] Step S233: Extract the perturbation feature sequence based on the multidimensional trend perturbation factor group and the induced perturbation factor group. Specifically, the multidimensional trend perturbation factor group (local volatility factor, trend slope mutation factor, range perturbation ratio factor, frequency domain perturbation energy factor) and the induced perturbation factor group (thermal mutation potential factor, performance drift index factor, probability perturbation memory factor) are concatenated to form a feature vector. The concatenated feature vector is standardized to ensure a consistent numerical range, for example, using Z-Score standardization. The standardized feature vectors of each sliding window are arranged in chronological order to form the perturbation feature sequence.

[0044] This approach, by extracting multidimensional trend disturbance factors and causal disturbance factors from historical power data, component temperature data, and inverter efficiency data, can comprehensively reflect the disturbance characteristics of the photovoltaic system, thereby improving the accuracy of abrupt change characteristic perception.

[0045] In one possible implementation, the mutation index tensor is used to output the prediction granularity. Step S250 further includes step S251, where the mutation index tensor is used as a tensor input, and the prediction probabilities of multiple candidate granularities are output based on the lightweight particle prediction network. Specifically, the lightweight particle prediction network is a neural network structure comprising an input layer, hidden layers, and an output layer. The input layer receives the mutation index tensor; for example, if the size of the mutation index tensor is 512×512×6, then the number of neurons in the input layer is 512×512×6. The hidden layer employs a lightweight design, such as using depthwise separable convolutions, which decompose standard convolutions into depthwise convolutions and pointwise convolutions, reducing computational cost. Assume the hidden layer has several depthwise separable convolutional layers, each followed by an activation function (such as ReLU) and a pooling layer (such as max pooling). The output layer outputs the prediction probabilities of multiple candidate granularities. Assuming there are three candidate granularities: 15 minutes, 30 minutes, and 1 hour, the output layer has 3 neurons. The softmax activation function is used to convert the output values ​​into a probability distribution. The training process of the lightweight particle prediction network is as follows: A large amount of historical mutation index tensor data and its corresponding actual optimal prediction granularity labels are collected as the training dataset; the cross-entropy loss function is used to measure the difference between the predicted probability distribution and the actual label distribution; stochastic gradient descent (SGD) or Adam optimization algorithm is used to backpropagate and update the network weights according to the loss function, iteratively training until the network converges.

[0046] Step S252: After filtering candidate granularities based on the predicted probabilities, the first predicted granularity is output. Specifically, a probability threshold (e.g., 0.5) is set, and the candidate granularity with the highest predicted probability is selected as the first predicted granularity. Alternatively, Top-k selection can be used, selecting the top k granularities with the highest probabilities for further evaluation to determine the first predicted granularity. Specifically, the candidate granularities are sorted in descending order of predicted probabilities. For example, for a predicted probability [0.6, 0.3, 0.1], the corresponding candidate granularities are sorted as 15 minutes (0.6), 30 minutes (0.3), and 1 hour (0.1). Based on the set k value, the top k candidate granularities with the highest probabilities are selected. For example, if k=2, the 15-minute and 30-minute granularities are selected. The selected top k candidate granularities are further evaluated, for example, based on factors such as historical prediction errors and computational resource consumption, to determine the optimal first predicted granularity. A specific evaluation method can be: for each selected candidate granularity, calculate its average prediction error over a past period, and select the granularity with the smallest average prediction error as the first predicted granularity.

[0047] Step S253: Optimize the global validation scale configuration by abstracting upwards from the first prediction granularity, and establish the second prediction granularity using the optimization result. Specifically, the first prediction granularity (e.g., 15 minutes) is abstracted upwards to a longer time scale, such as 30 minutes, 1 hour, 2 hours, etc. This can be achieved by merging the time periods of multiple adjacent first prediction granularities. For example, two 15-minute time periods are merged into one 30-minute time period, and four 15-minute time periods are merged into one 1-hour time period, etc. A genetic algorithm is used to optimize the global validation scale configuration. Individuals in the genetic algorithm are represented by a set of scale configuration parameters, and the fitness function can be defined as a comprehensive value of the stability and accuracy indices of the prediction results. For example, the fitness function is: Fitness = w1 × (1 − prediction error) + w2 × stability index, where w1 and w2 are weight coefficients, the prediction error can be calculated using indicators such as root mean square error (RMSE), and the stability index can be measured using indicators such as the standard deviation of the predicted values. Historical data is used to evaluate configurations at different scales. For each scale configuration, historical data is used for simulation prediction, and its prediction error and stability index are calculated to obtain the fitness value. Through iterative evolution using a genetic algorithm, including operations such as selection, crossover, and mutation, the optimal scale configuration with high fitness is gradually searched. Based on the optimal scale configuration obtained through optimization, a second prediction granularity is established.

[0048] This implementation employs a lightweight particle prediction network, which reduces computational complexity while maintaining prediction accuracy. The global optimization capability of the genetic algorithm ensures the stability and accuracy of the second prediction granularity.

[0049] In one possible implementation, the cloud map prediction data and the backtracking periodic dataset are sent to a mutation-aware scheduler to output the prediction granularity. The method further includes: determining the granularity reliability index of the prediction granularity; if the index determination result is a failure result, the redundant prediction compensation mechanism is activated, and redundant prediction is performed after reconstructing the prediction granularity.

[0050] Specifically, historical data from predictions made using the same prediction granularity over a past period (e.g., the past month) are collected, and the error between the predicted and actual values ​​is calculated, such as root mean square error (RMSE) and mean absolute error (MAE). A threshold for the granularity reliability index is set based on experience or experimentation; if the reliability index value for the current prediction granularity is lower than the threshold, the prediction is deemed unsuccessful.

[0051] When the prediction granularity fails to meet the reliability metric, a redundant prediction compensation mechanism is activated. Compensation strategies may include switching prediction models, adjusting prediction parameters, or increasing prediction frequency. Based on the strategy of the redundant prediction compensation mechanism, the prediction granularity is reconstructed. For example, if the original prediction granularity was 15 minutes, the compensation mechanism may adjust it to 30 minutes or a shorter time interval. The prediction process is then re-executed using the reconstructed prediction granularity to ensure the reliability of the prediction results.

[0052] This approach uses a granularity reliability index to promptly identify unreliable prediction granularities and activates a redundant prediction compensation mechanism, ensuring the reliability of prediction results and avoiding prediction errors caused by inappropriate prediction granularity selection.

[0053] Step S300: After the predictor matching activation of the multi-granularity prediction model cluster according to the prediction granularity, the heat map of mutation points is extracted according to the cloud map prediction data and the backtracking period dataset.

[0054] Specifically, a multi-granularity prediction model cluster comprises multiple models adapted to different time-granularity prediction needs, working collaboratively to improve the comprehensiveness and accuracy of predictions. For example, Long Short-Term Memory (LSTM) networks can be used for short-time-granularity (first prediction granularity) sequence prediction, while statistically based seasonal decomposition models are suitable for long-time-granularity (second prediction granularity) trend analysis. Based on the prediction granularity, a model matching algorithm is used to activate the corresponding predictor. Simultaneously, image processing and data analysis algorithms are used, combined with cloud image prediction data (analyzing cloud movement and shading) and backtesting periodic datasets (knowing historical power fluctuations, temperature change patterns, etc.), to extract abrupt change heatmaps. These heatmaps visually represent the location and extent of potential abrupt changes in photovoltaic power.

[0055] For example, for the first prediction granularity of 15 minutes, the LSTM model is activated to predict short-term fluctuations in photovoltaic power; for the second prediction granularity of 1 hour, the seasonal decomposition model is initiated to analyze the periodic trend of power. When extracting the heat map of abrupt change points, the movement trajectory of clouds in the cloud map is correlated with the layout map of photovoltaic modules. When changes in the position and density of clouds indicate that they may shade modules in a certain area, high-risk abrupt change points are marked at the corresponding positions on the heat map, with the intensity of the color indicating the degree of abrupt change.

[0056] In step S400, after the mutation attention of the activated predictor is aligned using the mutation point heatmap, photovoltaic power prediction is performed to establish a dual-granularity prediction result.

[0057] Specifically, the heatmap of abrupt change points is integrated into the activated predictor. By using an attention mechanism, the model focuses on features near the abrupt change points. When predicting photovoltaic power, the attention weights for different historical data and cloud map features are dynamically adjusted, enabling the model to accurately capture the impact of abrupt changes and generate prediction results that include power prediction values ​​at two granularities (first and second prediction granularities).

[0058] In one possible implementation, the mutation attention alignment of the activated predictor is performed using the mutation point heatmap. Step S400 further includes step S410, which involves thresholding the mutation point heatmap to extract a set of key mutation points. Specifically, the threshold can be determined using various methods, such as a fixed threshold set based on experience, or a dynamic threshold that can effectively distinguish between mutation points and non-mutation points by analyzing the histogram distribution of the mutation point heatmap. Each pixel in the mutation point heatmap is compared with the threshold; pixels greater than the threshold are considered key mutation points, while pixels less than or equal to the threshold are considered non-mutation points.

[0059] Step S420: Using the set of mutation key points, attention weights are assigned to the mutation values ​​mapped to each mutation key point, completing the initialization of the predictor mutation attention at the first prediction granularity. Specifically, for each mutation key point, attention weights are calculated based on its value (mutation intensity) in the mutation point heatmap. A normalization method can be used to convert the mutation intensity value into a weight value, and the calculated attention weights are mapped to the corresponding positions in the predictor to initialize the predictor mutation attention at the first prediction granularity.

[0060] Step S430 involves assigning gradual attention weights to each mutation point in the mutation keypoint set, completing the predictor mutation attention initialization at the second prediction granularity. Specifically, the gradual attention weight assignment can be implemented by defining a gradual function that progressively changes the attention weights over time. For example, a linear gradual function can be used to gradually decrease or increase the attention weights over time; alternatively, a non-linear gradual function, such as a Gaussian function, can be used to concentrate the attention weights around a certain keypoint. The gradual function is then applied to the mutation keypoint set to assign a gradual attention weight to each mutation point.

[0061] This implementation extracts a set of key abrupt change points through threshold segmentation, focusing on the regions most likely to trigger photovoltaic power abrupt changes, reducing invalid computations and improving prediction efficiency. In short-term predictions, attention weighting based on abrupt change values ​​allows the predictor to focus on key points with high abrupt change intensity, accurately capturing short-term power changes. In long-term predictions, gradual attention weighting considers the gradual impact of abrupt change key points over time, enabling the predictor to better grasp power change trends.

[0062] Step S500: After performing self-verification correction on the dual-granularity prediction results, output the photovoltaic power prediction results.

[0063] Specifically, the prediction results at the first prediction granularity (short-term) and the second prediction granularity (long-term) are compared to identify differences and discrepancies between the short-term predictions and the long-term trend. The consistency between the two prediction granularities is assessed to determine if the short-term predictions align with the long-term trend and whether correction is necessary. Based on the consistency assessment, if significant deviations are found in the short-term predictions, the corresponding correction mechanism is activated to revise the short-term predictions to better reflect the long-term trend. The corrected short-term predictions are then output as the final photovoltaic power prediction, ensuring that the predictions retain both short-term details and long-term trend accuracy.

[0064] In one possible implementation, after self-verification correction of the dual-granularity prediction results, the photovoltaic power prediction results are output. Step S500 further includes step S510, which involves interpolating and aligning the dual-granularity prediction results to construct a prediction value difference tensor on the same time axis. Specifically, assume the first prediction granularity is 15 minutes and the second prediction granularity is 1 hour. For the result of the second prediction granularity, linear interpolation is used to refine it to a time interval of 15 minutes. After interpolation, both prediction results have the same time interval, thus enabling comparison on the same time axis. At each time point, the difference between the predicted values ​​of the first and second prediction granularities is calculated. For example, at time point t1, the first prediction granularity value is 95kW, and the interpolated value of the second prediction granularity is 100kW, then the difference is -5kW. These difference values ​​are organized into a tensor structure containing a time dimension and a difference value dimension. For example, the difference tensor can be a two-dimensional table, where rows represent time points and columns represent corresponding difference values.

[0065] Step S520: Calculate the consistency coefficient and offset drift index using the predicted value difference tensor to generate a consistency score for the granularity results. Specifically, calculate the Pearson correlation coefficient between the predicted values ​​of the first and second predicted granularities. Other consistency indices, such as the consistency slope, can also be calculated to obtain the consistency coefficient. The offset drift index can be the average deviation or the trend deviation. The average deviation is obtained by averaging all the difference values ​​in the predicted value difference tensor. For example, an average deviation of -2kW indicates that the first predicted granularity is generally 2kW lower than the second predicted granularity. The trend deviation is obtained by analyzing the trend of the difference values ​​over time using methods such as linear regression. For example, a trend deviation of 0.5kW / hour indicates that the deviation of the first predicted granularity relative to the second predicted granularity increases by 0.5kW per hour.

[0066] A comprehensive consistency score is generated based on the consistency coefficient and the deviation drift index. For example, the scoring formula is set as: Consistency Score = α1 × Consistency Coefficient + α2 × (1 - |Average Deviation| / Maximum Deviation), where α1 and α2 are weighting coefficients.

[0067] Step S530: If the consistency score of the granularity result meets the activation condition, the residual correction network is activated. Specifically, a threshold for the consistency score is preset. If the consistency score is lower than the threshold, it is considered that the residual correction network needs to be activated for further correction. The residual correction network is a pre-trained neural network model used to correct the first prediction granularity result based on the predicted value difference tensor and the enhancement factor.

[0068] Step S540: The residual correction network corrects the first prediction granularity result based on the predicted value difference tensor and the enhancement factor, and outputs the result as the photovoltaic power prediction result. Specifically, the predicted value difference tensor constructed in step S510 is used as one of the input features, and other factors that may affect the prediction accuracy, such as historical prediction errors and environmental factors (temperature, light intensity, etc.), are collected as enhancement factors. For example, the enhancement factor includes the average prediction error over the past 24 hours and the current temperature. The residual correction network calculates the correction value for the first prediction granularity result based on the input predicted value difference tensor and the enhancement factor, and outputs the correction value. The correction value is added to the first prediction granularity result to obtain the final photovoltaic power prediction result.

[0069] This approach accurately identifies and quantifies the differences between two prediction results by interpolating and aligning the two-granularity predictions and performing difference analysis. By using a residual correction network combined with an enhancement factor to correct the first prediction granularity result, prediction errors can be effectively reduced, improving the accuracy of photovoltaic power prediction.

[0070] In one possible implementation, after outputting the photovoltaic power prediction result, the method further includes: establishing a verification time node based on the first prediction granularity; reading photovoltaic power at the verification time node to establish verification data; using the verification data and the photovoltaic power prediction result to perform deviation verification and establish deviation feedback; and using the deviation feedback as correction data to perform adaptive correction compensation for the prediction.

[0071] Specifically, assuming the initial prediction granularity is 15 minutes, a verification time node is established every 15 minutes based on the prediction start time for subsequent verification operations. For example, starting at 8:00 AM, the verification time nodes are 8:15, 8:30, 8:45, and so on. At each verification time node, the actual photovoltaic power output value is read in real time by power sensors installed in the photovoltaic power station. The read actual power value is recorded to form a verification dataset. For example, if the actual power is read as 98.5 kW at the 8:15 node, this data is recorded.

[0072] The verification data is compared point-by-point with the photovoltaic power prediction results, and the deviation value at each verification time point is calculated. For example, if the predicted value is 100kW and the actual value is 98.5kW, the deviation is -1.5kW. Overall deviation statistics, such as mean absolute deviation and root mean square deviation, are calculated. A deviation feedback data structure is constructed that includes deviation values, deviation statistics, and verification time point information.

[0073] Choose a correction algorithm, such as a bias compensation algorithm, which directly subtracts the corresponding bias value from subsequent prediction results; or use an adaptive filtering algorithm, such as Kalman filtering, to dynamically correct the prediction results. Initialize the relevant parameters of the correction algorithm based on existing bias feedback data. For example, in Kalman filtering, initialize the state estimate and the error covariance matrix. During new prediction processes, use stored bias feedback data to correct and compensate for the prediction results in real time. For example, if the predicted value is 102kW, and based on historical biases, the correction algorithm calculates a correction value of -1.0kW, then the corrected prediction output is 101kW. As more bias feedback data accumulates, continuously update the parameters of the correction algorithm to achieve adaptive optimization of the prediction model and continuously improve prediction accuracy.

[0074] This approach, by reading actual power data in real time and calculating deviations at verification points, can promptly identify discrepancies between predicted and actual values. Using this deviation data for adaptive correction and compensation can continuously improve the accuracy of photovoltaic power prediction and reduce the impact of prediction errors on grid dispatch and power plant operation.

[0075] In one possible implementation, the deviation feedback is used as correction data to perform adaptive correction compensation for prediction. The method further includes: performing continuous feedback monitoring of the deviation feedback, determining whether the continuous feedback monitoring result meets a trigger threshold; if the continuous feedback monitoring result meets the trigger threshold, then triggering an online incremental correction mechanism to perform prediction training reconstruction.

[0076] Specifically, a real-time monitoring system is established to continuously track and record deviation feedback data, including the deviation value and changes in deviation statistics at each verification time point. The monitored data is dynamically analyzed to calculate real-time deviation statistics, such as real-time average deviation and standard deviation, and their trends are observed. Trigger thresholds are set based on the actual operating experience of photovoltaic power plants and the required prediction accuracy. For example, a prediction deviation exceeding 5kW or a standard deviation exceeding 3kW is considered significant, requiring the triggering of an online incremental correction mechanism. Other factors, such as grid dispatch requirements and power plant operating status, can also be considered to comprehensively set trigger thresholds. During continuous feedback monitoring, the calculated deviation statistics are compared with the trigger thresholds in real time.

[0077] When the triggering conditions are met, the online incremental correction mechanism is automatically activated, initiating the prediction training reconstruction process. The latest deviation feedback data and relevant input data, such as historical power data, weather data, and environmental parameters, are collected to prepare for retraining the prediction model. Online learning algorithms, such as incremental learning neural networks and online sequential learning Support Vector Machines, are used to update and train the prediction model using the newly collected data. For example, for a neural network model, a forward and backward propagation is performed using new data to update the model's weight parameters. During training, the model's performance is verified in real time, such as through cross-validation, to ensure the model's accuracy and generalization ability. Based on the verification results, the model is optimized and adjusted, such as by adjusting hyperparameters. After training reconstruction is completed, a strategy is determined to either directly replace the original prediction model or integrate the new model with the old model, for example, by using model fusion techniques to improve prediction performance.

[0078] This approach, through continuous feedback monitoring and online incremental correction mechanisms, enables the prediction model to adapt in a timely manner to changes in the operating environment and equipment status of photovoltaic power plants, as well as fluctuations in external conditions (such as weather), thus maintaining the high accuracy and reliability of the prediction model.

[0079] This application's embodiments employ techniques such as collecting historical power, component temperature, and inverter efficiency data based on the current time node to construct a backtracking periodic dataset, and simultaneously reading cloud map prediction data. These two types of data are input into a mutation-aware scheduler to generate first and second prediction granularities. Based on the prediction granularity, a matching predictor is activated, mutation point heatmaps are extracted and attention alignment is performed, photovoltaic power prediction is executed to obtain dual-granularity results, and the final prediction result is output through self-verification correction. These techniques solve the technical problem that existing distributed photovoltaic power prediction cannot meet the grid's dual requirements for prediction accuracy and flexibility, achieving the technical effect of improving the accuracy and flexibility of distributed photovoltaic power prediction.

[0080] In the above text, refer to Figure 1 A distributed photovoltaic power prediction method based on adaptive learning according to an embodiment of the present invention is described in detail. Next, reference will be made to... Figure 2 This invention describes a distributed photovoltaic power prediction system based on adaptive learning according to an embodiment of the present invention.

[0081] An adaptive learning-based distributed photovoltaic power prediction system according to an embodiment of the present invention addresses the technical problem that existing distributed photovoltaic power prediction systems cannot meet the power grid's dual requirements for prediction accuracy and flexibility, thereby improving the accuracy and flexibility of distributed photovoltaic power prediction. The adaptive learning-based distributed photovoltaic power prediction system includes: a historical data acquisition module 10, a granular prediction module 20, a change-point heatmap extraction module 30, a photovoltaic power prediction module 40, and a self-verification correction module 50.

[0082] The historical data acquisition module 10 is used to collect historical power data, component temperature data, and inverter efficiency data within a preset time period, using the current time node as the zero point, and establish a backtracking period dataset. The granularity prediction module 20 is used to read the cloud map prediction data, send the cloud map prediction data and the backtracking period dataset to the mutation sensing scheduler, and output the prediction granularity. The prediction granularity includes a first prediction granularity and a second prediction granularity. The second prediction granularity is constructed based on the first prediction granularity and is larger than the first prediction granularity. The mutation point heatmap extraction module 30 is used to extract the mutation point heatmap based on the cloud map prediction data and the backtracking period dataset after matching and activating the predictors of the multi-granularity prediction model cluster according to the prediction granularity. The photovoltaic power prediction module 40 is used to perform photovoltaic power prediction after aligning the mutation attention of the activated predictors using the mutation point heatmap and establish a dual-granularity prediction result. The self-verification correction module 50 is used to perform self-verification correction on the dual-granularity prediction result and output the photovoltaic power prediction result.

[0083] The specific configuration of the granularity prediction module 20 will be described in detail below. As mentioned above, the cloud image prediction data and the backtracking period dataset are sent to the mutation-aware scheduler to output the prediction granularity. The granularity prediction module 20 may further include: a spatial feature extraction unit for extracting spatial features from the cloud image prediction data after temporal sorting, using a convolutional image processing algorithm. The spatial features include cloud edge density gradient map, cloud edge displacement velocity vector, and temporal change rate of the occlusion coverage area; a local occlusion mutation factor map establishment unit for establishing a local occlusion mutation factor map using the spatial features; a temporal dimension mutation analysis unit for performing temporal dimension mutation analysis based on a sliding window on the backtracking period dataset to extract the perturbation feature sequence; a mutation exponent tensor establishment unit for establishing a mutation exponent tensor after standardization processing based on the perturbation feature sequence and the local occlusion mutation factor map, performing feature splicing and tensor fusion; and a prediction granularity output unit for outputting the prediction granularity using the mutation exponent tensor.

[0084] Specifically, the backtracking period dataset undergoes time-dimensional mutation analysis based on a sliding window to extract disturbance feature sequences. The time-dimensional mutation analysis unit may further include: a direct volatility feature extraction subunit for extracting direct volatility features within the sliding window and establishing a multi-dimensional trend disturbance factor group, which includes a local volatility factor, a trend slope mutation factor, a range disturbance ratio factor, and a frequency domain disturbance energy factor; an induced disturbance feature extraction subunit for extracting induced disturbance features within the sliding window and establishing an induced disturbance factor group, which includes a thermal mutation potential factor, a performance drift index factor, and a probability disturbance memory factor; and a disturbance feature sequence extraction subunit for extracting disturbance feature sequences based on the multi-dimensional trend disturbance factor group and the induced disturbance factor group.

[0085] The prediction granularity output unit, which uses the mutation exponent tensor to output the prediction granularity, may further include: a particle prediction subunit for taking the mutation exponent tensor as a tensor input and outputting prediction probabilities of multiple candidate granularities based on a lightweight particle prediction network; a candidate granularity filtering subunit for filtering candidate granularities according to the prediction probabilities and outputting the first prediction granularity; and a global verification scale configuration optimization subunit for performing upward abstraction of the first prediction granularity to optimize the global verification scale configuration and establishing a second prediction granularity using the configuration optimization results.

[0086] The specific configuration of the self-verification correction module 50 will be described in detail below. As mentioned above, after performing self-verification correction on the dual-granularity prediction results, the photovoltaic power prediction results are output. The self-verification correction module 50 may further include: an interpolation alignment unit for interpolating and aligning the dual-granularity prediction results to construct a prediction value difference tensor on the same time axis; a granularity result consistency score generation unit for calculating the consistency coefficient and offset drift index using the prediction value difference tensor to generate a granularity result consistency score; a residual correction network activation unit for activating the residual correction network if the granularity result consistency score meets the activation condition; and a correction unit for correcting the first prediction granularity result using the residual correction network based on the prediction value difference tensor and the enhancement factor, and then outputting it as the photovoltaic power prediction result.

[0087] The specific configuration of the photovoltaic power prediction module 40 will be described in detail below. As mentioned above, the photovoltaic power prediction module 40, which uses the mutation point heatmap for activated predictor mutation attention alignment, may further include: a threshold segmentation unit for performing threshold segmentation on the mutation point heatmap to extract a set of mutation key points; a mutation value attention weighting unit for using the set of mutation key points to assign mutation value attention weights to each mutation key point, completing the predictor mutation attention initialization at the first prediction granularity; and a gradual attention weighting unit for assigning gradual attention weights to each mutation point in the set of mutation key points, completing the predictor mutation attention initialization at the second prediction granularity.

[0088] The system may further include, after outputting the photovoltaic power prediction result: a verification time node establishment module for establishing a verification time node based on the first prediction granularity; a verification data establishment module for reading photovoltaic power at the verification time node and establishing verification data; a deviation verification module for using the verification data and the photovoltaic power prediction result to perform deviation verification and establish deviation feedback; and an adaptive correction compensation module for using the deviation feedback as correction data to perform adaptive correction compensation for the prediction.

[0089] The system further includes: a continuous feedback monitoring module for performing continuous feedback monitoring of the deviation feedback and determining whether the continuous feedback monitoring result meets the trigger threshold; and a prediction training reconstruction module for triggering an online incremental correction mechanism and performing prediction training reconstruction if the continuous feedback monitoring result meets the trigger threshold.

[0090] The system further includes: a granularity reliability determination module for determining the granularity reliability index of the prediction granularity; and a redundant prediction compensation module for activating the redundant prediction compensation mechanism and reconstructing the prediction granularity to perform redundant prediction if the index determination result is a failure result.

[0091] The distributed photovoltaic power prediction system based on adaptive learning provided in this embodiment of the invention can execute the distributed photovoltaic power prediction method based on adaptive learning provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method.

[0092] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0093] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A distributed photovoltaic power prediction method based on adaptive learning, characterized in that, The method includes: Using the current time node as the zero point, perform data collection of historical power data, component temperature data, and inverter efficiency data within a preset time period to establish a backtracking period dataset. After reading the cloud map prediction data, the cloud map prediction data and the backtracking periodic dataset are sent to the mutation-aware scheduler to output the prediction granularity. The prediction granularity includes a first prediction granularity and a second prediction granularity. The second prediction granularity is constructed based on the first prediction granularity and is larger than the first prediction granularity. After the predictor matching activation of the multi-granularity prediction model cluster according to the prediction granularity, the heat map of mutation points is extracted according to the cloud map prediction data and the backtracking period dataset. After the mutation attention of the activated predictor is aligned using the mutation point heatmap, photovoltaic power prediction is performed to establish a dual-granularity prediction result. After performing self-verification correction on the dual-granularity prediction results, the photovoltaic power prediction results are output.

2. The distributed photovoltaic power prediction method based on adaptive learning as described in claim 1, characterized in that, The cloud map prediction data and the backtracking periodic dataset are sent to the mutation-aware scheduler, and the prediction granularity is output, including: After the mutation-aware scheduler receives the cloud map prediction data and the backtracking period data, it performs mutation feature awareness as follows: After sorting the cloud image prediction data by time, spatial features are extracted by convolutional image processing algorithm. The spatial features include cloud edge density gradient map, cloud edge displacement velocity vector, and time change rate of occlusion coverage area. A local occlusion mutation factor map was constructed using the aforementioned spatial features; The backtracking period dataset is subjected to time-dimensional mutation analysis based on a sliding window to extract perturbation feature sequences. After standardization of the perturbation feature sequence and the local occlusion mutation factor map, feature splicing and tensor fusion are performed to establish the mutation index tensor. The prediction granularity is output using the mutation exponent tensor.

3. The distributed photovoltaic power prediction method based on adaptive learning as described in claim 2, characterized in that, The backtracking period dataset is subjected to time-dimensional mutation analysis based on a sliding window to extract perturbation feature sequences, including: Direct volatility features are extracted within the sliding window to establish a multidimensional trend disturbance factor group, which includes local volatility factor, trend slope abrupt change factor, range disturbance ratio factor, and frequency domain disturbance energy factor. Within the sliding window, induced disturbance features are extracted, and an induced disturbance factor group is established, which includes a thermal mutation potential factor, a performance drift index factor, and a probability disturbance memory factor. The perturbation feature sequence is extracted based on the multidimensional trend perturbation factor group and the induced perturbation factor group.

4. The distributed photovoltaic power prediction method based on adaptive learning as described in claim 2, characterized in that, The prediction granularity is output using the mutation exponential tensor, including: Using the mutation exponent tensor as tensor input, a lightweight particle prediction network outputs prediction probabilities at multiple candidate granularities. After filtering candidates by granularity based on the predicted probability, the first predicted granularity is output. The global verification scale configuration is optimized by abstracting upwards from the first prediction granularity, and the second prediction granularity is established using the configuration optimization result.

5. The distributed photovoltaic power prediction method based on adaptive learning as described in claim 1, characterized in that, After performing self-verification correction on the dual-granularity prediction results, the photovoltaic power prediction results are output, including: The dual-granularity prediction results are interpolated and aligned to construct a prediction difference tensor on the same time axis; The consistency coefficient and offset drift index are calculated using the predicted value difference tensor to generate a consistency score for granular results. If the consistency score of the granularity result meets the activation condition, then the residual correction network is activated; The residual correction network is used to correct the first prediction granularity result based on the prediction difference tensor and the enhancement factor, and then the result is output as the photovoltaic power prediction result.

6. The distributed photovoltaic power prediction method based on adaptive learning as described in claim 1, characterized in that, Mutation attention alignment of the activated predictor using the mutation point heatmap includes: Threshold segmentation is performed on the heatmap of the mutation points to extract the set of key mutation points; The mutation value attention of each mutation key point is assigned using the set of mutation key points to complete the mutation attention initialization of the predictor at the first prediction granularity. The set of mutation key points is weighted with a gradual attention value for each mutation point to complete the mutation attention initialization of the predictor at the second prediction granularity.

7. The distributed photovoltaic power prediction method based on adaptive learning as described in claim 1, characterized in that, After outputting the photovoltaic power prediction results, the following are included: Establish verification time nodes based on the first prediction granularity; Photovoltaic power is read at the specified verification time point to establish verification data; Using the verification data and the photovoltaic power prediction results, deviation verification is performed, and deviation feedback is established; The deviation feedback is used as correction data to perform adaptive correction compensation for prediction.

8. The distributed photovoltaic power prediction method based on adaptive learning as described in claim 7, characterized in that, Using the deviation feedback as correction data to perform adaptive correction compensation for prediction further includes: Perform continuous feedback monitoring of deviation feedback and determine whether the continuous feedback monitoring results meet the trigger threshold; If the continuous feedback monitoring results meet the trigger threshold, the online incremental correction mechanism is triggered to perform prediction training reconstruction.

9. The distributed photovoltaic power prediction method based on adaptive learning as described in claim 1, characterized in that, Sending the cloud map prediction data and the backtracking periodic dataset to the mutation-aware scheduler, outputting the prediction granularity, and also including: The predicted granularity is then assessed using a granularity reliability index. If the indicator is judged as failing, the redundant prediction compensation mechanism is activated, and the redundant prediction is executed after the prediction granularity is reconstructed.

10. A distributed photovoltaic power prediction system based on adaptive learning, characterized in that, The system is used to implement the distributed photovoltaic power prediction method based on adaptive learning as described in any one of claims 1-9, and the system comprises: The historical data acquisition module is used to collect historical power data, component temperature data, and inverter efficiency data within a preset time period, with the current time node as the zero point, and to establish a backtracking period dataset. The granularity prediction module is used to read the cloud map prediction data, send the cloud map prediction data and the backtracking periodic dataset to the mutation-aware scheduler, and output the prediction granularity. The prediction granularity includes a first prediction granularity and a second prediction granularity. The second prediction granularity is constructed based on the first prediction granularity and the second prediction granularity is greater than the first prediction granularity. The mutation point heatmap extraction module is used to extract mutation point heatmaps based on the cloud map prediction data and the backtracking period dataset after the predictor matching activation of the multi-granularity prediction model cluster according to the prediction granularity. The photovoltaic power prediction module is used to perform photovoltaic power prediction after the predictor is activated by the mutation attention alignment using the mutation point heatmap, and to establish a dual-granularity prediction result. The self-verification correction module is used to perform self-verification correction on the dual-granularity prediction results and then output the photovoltaic power prediction results.

Images