Distributed photovoltaic power generation prediction method and system
By constructing a multi-source high-quality data processing pipeline, performing anomaly identification and repair, extracting multi-dimensional features, and using machine learning algorithms for multi-timescale rolling prediction, the problems of low accuracy and poor robustness in distributed photovoltaic power generation prediction are solved, and grid operation support is achieved in high-penetration distributed photovoltaic access scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- STATE GRID ZHEJIANG ELECTRIC POWER CO LTD NINGBO POWER SUPPLY CO
- Filing Date
- 2025-12-05
- Publication Date
- 2026-04-10
AI Technical Summary
Existing technologies for distributed photovoltaic (PV) power generation forecasting suffer from problems such as single data source, crude feature engineering, weak abnormal data processing, single prediction dimension, and fragmented time scale. These issues result in low prediction accuracy, poor robustness, and difficulty in quantifying uncertainty, making it difficult to support grid operation in scenarios with high penetration of distributed PV.
By constructing a multi-source high-quality data processing pipeline, anomaly identification and repair are performed, multi-dimensional features are extracted, and machine learning algorithms are used for multi-timescale rolling prediction, outputting probability interval prediction results, thus forming an end-to-end prediction pipeline.
It significantly improves the accuracy and robustness of distributed photovoltaic power generation forecasting, provides probabilistic outputs at multiple time scales, supports risk assessment and decision-making by grid dispatchers, and enhances the support capabilities for grid operation.
Smart Images
Figure CN121840563A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of photovoltaic technology, and in particular to a method and system for predicting distributed photovoltaic power generation. Background Technology
[0002] With distributed photovoltaic (PV) power generation emerging as a crucial component of clean energy, it has experienced explosive growth in my country's power distribution network in recent years. The large-scale integration of distributed PV power sources into distribution substations and feeder levels has significantly altered the power flow distribution and operational characteristics of traditional distribution networks. However, PV power generation inherently possesses intermittency, volatility, and strong weather dependence. Its output is affected by various environmental factors such as solar radiation intensity, cloud cover, temperature, and humidity, making it difficult to maintain stability and control. When distributed PV penetration is high, the lack of accurate prediction of its future output will pose severe challenges to critical operations such as load balancing, dispatching plans, maintenance schedules, and voltage control of the power grid, and may even threaten the safe and stable operation of the system.
[0003] To address these issues, the power system field has conducted research on various photovoltaic power generation prediction technologies. Early methods were mainly based on physical models, such as using solar position algorithms combined with atmospheric radiative transfer models to calculate the theoretical maximum output, and then correcting it with equipment efficiency. While these methods have a certain degree of physical interpretability, they are slow to respond to local micro-meteorological conditions (such as sudden cloud cover) and have difficulty adapting to the individual differences of massive distributed units, resulting in limited prediction accuracy.
[0004] Currently, statistical learning methods are gradually becoming mainstream, including time series models (such as ARIMA), support vector machines (SVM), and shallow neural networks. These methods improve accuracy to some extent by mining the temporal patterns of historical power generation data for prediction. However, current solutions generally suffer from the following key drawbacks: Single data source: Most models rely solely on historical power generation data and fail to effectively integrate high spatiotemporal resolution gridded meteorological forecast information, resulting in insufficient perception of external disturbances such as sudden weather changes. The feature engineering is crude: the input features are usually just raw power values or simple lag terms, lacking a deep characterization of the weather-power generation coupling relationship, and failing to systematically introduce time context (such as seasons, holidays) or derived features (such as clear sky index, power change rate), which limits the model's expressive power. Weak handling of abnormal data: Actual collected distributed photovoltaic data often contains anomalies such as missing data, jumps, and freezes. However, existing methods often directly use raw data for modeling without establishing a sound quality control mechanism, which seriously affects the quality of training samples. Single prediction dimension: Most systems only provide point prediction results, which cannot quantify prediction uncertainty and are difficult to support schedulers in risk assessment and robust decision-making; Time scale fragmentation: Ultra-short-term (hourly), short-term (daily), and medium-term (multi-day) forecasts are usually implemented independently by different systems, lacking rolling updates and collaborative optimization under a unified architecture.
[0005] In summary, existing technical solutions lack a complete end-to-end, engineerable prediction pipeline, making it difficult to achieve highly reliable, multi-scale, and probabilistic distributed photovoltaic prediction services in real power grid environments. Based on this, this proposal is put forward. Summary of the Invention
[0006] The purpose of this invention is to provide a distributed photovoltaic power generation prediction method and system that integrates high-quality data from multiple sources, has the ability to construct fine features, supports rolling prediction at multiple time scales and outputs probability intervals, so as to meet the core requirements of new power systems for observable, measurable and controllable new energy sources.
[0007] To achieve the above objectives, the technical solution of the present invention is as follows: A method for predicting distributed photovoltaic power generation includes the following steps: S1. Obtain historical power generation data of distributed photovoltaics and meteorological data of corresponding spatiotemporal locations, and perform time alignment on the acquired data to form a multidimensional time-series dataset; S2. Perform anomaly identification and repair processing on the multidimensional time-series dataset to generate training data; S3. Based on the training data, the historical power generation data and the meteorological data at the corresponding time are aligned by timestamp, and the historical power generation lag features, the meteorological features of the same period and the time context features are extracted based on the preset time window, and combined to form an initial feature set. Then, the feature construction algorithm is used to select the feature subset of the initial feature set to form the optimal feature subset. S4. Using the aforementioned optimal feature subset, a machine learning algorithm is employed to train a prediction model; S5. Based on the trained prediction model, perform multi-timescale rolling prediction and output the predicted point values and corresponding confidence level probability interval prediction results within the preset future time period. S6. Publish the probability interval prediction results to the outside world for use by the scheduling and management system.
[0008] Furthermore, in S1, the meteorological data includes gridded irradiance, temperature, and humidity data, with a data time interval of 15 minutes; the historical power generation data comes from the electricity consumption information collection system and is aggregated by transformer area or feeder.
[0009] Furthermore, the anomaly identification and repair process in S2 includes one or more of the following: Determine if there are missing values; if so, fill them using interpolation or the mean of nearby time periods. Alternatively, it can determine whether there are any abnormal mutations, and if so, make corrections based on sliding window trend detection; Alternatively, it can determine whether there are continuous constant value anomalies. If they exist and the duration exceeds the threshold, it can be reconstructed based on the daily cycle pattern or meteorological changes. Alternatively, it can determine whether the power generation time exceeds the reasonable sunshine period; if it does, it should be marked as invalid and filled in. Or determine whether there is an abnormal power generation trend; if so, fill in the data based on the power change pattern of normal days in the same historical period. Alternatively, it can determine whether there are upper or lower boundary anomalies in the meteorological data. If so, it can truncate or fill the data with nearby valid values based on the preset threshold range of the meteorological data.
[0010] Furthermore, in S3, the preset time window includes several past time steps; The historical power generation lag characteristics include the power generation values at several time points prior to the current moment; The meteorological characteristics mentioned above include the forecast irradiance, temperature, and humidity at the current moment and several future time points; The time context features include one or more of the following: hour identifier, weekday type, seasonal factor, or whether it is a holiday.
[0011] Furthermore, the initial feature set also includes derived features, which include one or more of the following: First-order difference or rate of change of power generation; Or the clear sky index, which is the ratio of measured irradiance to theoretical maximum irradiance; Or the moving average or trend slope of meteorological data.
[0012] Furthermore, the feature construction algorithm in S3 includes at least one of recursive feature elimination, feature importance ranking based on a tree model, or principal component analysis.
[0013] Furthermore, in S3, the feature subset selection includes the following process: S31. Assign scores or weights to each feature in the initial feature set using the feature construction algorithm; S32. Sort the scores from high to low, and select the top k features in sequence to form a candidate subset, where k can take multiple values to form multiple candidate subsets; S33. Input multiple candidate subsets into the validation model for cross-validation and calculate the corresponding prediction error index. The prediction error index includes one or more of the following: root mean square error, mean absolute percentage error, or continuous ranking probability score. S34. Select the candidate subset that minimizes the prediction error index as the optimal feature subset.
[0014] Furthermore, in S5, the multiple time scales include: Ultra-short term, meaning the next 4 hours; Short term, meaning the next 3 days; The medium term refers to the next 7 days.
[0015] Furthermore, the probability interval prediction results include at least the upper and lower bounds of the prediction intervals corresponding to the three confidence levels of 85%, 90%, and 95%.
[0016] A distributed photovoltaic power generation prediction system based on the above method includes: The data acquisition module is used to acquire historical power generation data of distributed photovoltaic power and meteorological data from the electricity consumption information collection system and the meteorological system, respectively. The data quality control module is used to identify and repair anomalies in the acquired data and generate training data. The feature engineering module is used to build and optimize the feature set for training the prediction model, forming the optimal feature subset. The model training module is used to train the prediction model based on the optimal feature subset; The prediction execution module is used to call the trained prediction model to perform rolling predictions at multiple time scales and output point prediction values and probability interval prediction results at different confidence levels. The results publishing module is used to provide services to external parties with the probability interval prediction results.
[0017] The advantages of this invention are: 1. By constructing an end-to-end prediction pipeline of "high-quality data - intelligent features - multi-scale model - probabilistic output", the core problems of low prediction accuracy, poor robustness and difficulty in quantifying uncertainty in existing technologies are effectively solved, and the grid operation support capability in high-penetration distributed photovoltaic access scenarios is significantly improved.
[0018] 2. In terms of data quality control, in addition to handling routine missing values, abrupt changes, and continuous constant values, we also introduced a power generation trend anomaly identification and meteorological data upper and lower bound anomaly verification mechanism to fully ensure the physical rationality and temporal consistency of the input data, laying a high-quality data foundation for subsequent modeling.
[0019] 3. In terms of feature construction, historical power generation lag terms, concurrent and forecast meteorological elements, time context (such as hours, seasons, holidays) and derived features (such as power change rate, clear sky index, meteorological moving average and trend slope) are spatiotemporally aligned and fused to form an initial feature set. A feature subset selection mechanism based on cross-validation is adopted to automatically select the optimal feature combination with the smallest prediction error from multiple candidate subsets, balancing model accuracy and complexity, and significantly improving feature expression ability.
[0020] 4. In terms of forecast output, it supports rolling forecasts at multiple time scales, including ultra-short-term, short-term, and medium-term, and outputs the upper and lower bounds of probability intervals at various confidence levels such as 85%, 90%, and 95%. This probabilistic result provides dispatchers with a basis for risk quantification and can be flexibly applied to different business scenarios such as reserve capacity configuration, maintenance plan optimization, and safety margin assessment. Attached Figure Description
[0021] Figure 1 This is a schematic diagram of the distributed photovoltaic power generation prediction function architecture in the embodiment; Figure 2 This is a schematic diagram of the distributed photovoltaic power generation point prediction process in the embodiment; Figure 3 This is a schematic diagram of the distributed photovoltaic power generation data processing flow in the embodiment; Figure 4 This is a schematic diagram of the meteorological data processing flow in the embodiment. Detailed Implementation
[0022] The present invention will be further described in detail below with reference to embodiments. It should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., used in this document indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting the present invention.
[0023] This embodiment proposes a distributed photovoltaic (PV) power generation prediction method. By constructing an end-to-end prediction pipeline of "high-quality data - intelligent features - multi-scale model - probabilistic output," it effectively solves the core problems of low prediction accuracy, poor robustness, and difficulty in quantifying uncertainty in existing technologies, significantly improving the grid operation support capability for high-penetration distributed PV access scenarios. Figure 1 and Figure 2 As shown, the process includes the following steps.
[0024] Step S1. Acquire historical power generation data of distributed photovoltaic systems and corresponding spatiotemporal meteorological data, and perform time alignment on the acquired data to form a multidimensional time-series dataset. In this step, the meteorological data includes gridded irradiance, temperature, and humidity data, with a data time interval of 15 minutes; the historical power generation data comes from the electricity consumption information collection system and is aggregated by transformer area or feeder.
[0025] Since power generation data and meteorological data may come from different systems and have slight time offsets (e.g., power generation is recorded at 08:02, while meteorological data is recorded at 08:00), time alignment is required. This can be achieved by precisely matching timestamps (e.g., rounding to the nearest 15 minutes) or by interpolation to ensure that the power generation value P(t) at each time t corresponds strictly to the meteorological values G(t), T(t), etc.
[0026] Step S2. Perform anomaly identification and repair processing on the multidimensional time-series dataset to generate training data. For example... Figure 3 and Figure 4 As shown, anomaly identification and repair processes include the following six types: 1. Determine if missing values exist; if so, fill them using interpolation or the mean of nearby time periods. The system first checks if there are missing values in the multidimensional time series dataset (such as missing power or meteorological data due to communication interruption or equipment failure). If missing values are found, an appropriate filling strategy will be selected based on the data characteristics: for short-term, isolated missing points, linear interpolation or spline interpolation will be used, and smooth estimation will be performed using the effective observations before and after; for continuous missing values or boundary missing values, the mean of effective data from nearby time periods (such as the historical mean of the previous time period, the next time period, or the same time period) will be used to fill the missing values, so as to maintain the continuity and physical rationality of the time series.
[0027] 2. Determine if any abrupt changes or anomalies exist, and if so, make corrections based on sliding window trend detection. Abrupt changes or anomalies refer to drastic jumps in power generation within a short period (e.g., between adjacent 15-minute intervals) that defy physical laws, such as a sudden surge from near 0 kW to over 90% of rated power, or vice versa. Such changes are typically caused by communication failures, momentary sensor malfunctions, or data acquisition errors. Sliding window trend detection identifies abrupt changes or anomalies that violate physical laws by analyzing the rate of power change or the degree of fluctuation within a local time segment, and then corrects the data accordingly.
[0028] 3. Determine if there are any continuous constant value anomalies. If they exist and the duration exceeds the threshold, reconstruct the data based on the daily cycle pattern or meteorological changes.
[0029] In the actual data collection process of distributed photovoltaic (PV) power generation, power records often remain completely unchanged for a period of time (e.g., the same value for more than 2 consecutive hours, such as 300 kW) due to communication interruptions, inverter failures, or data freezing. This phenomenon is clearly inconsistent with the physical characteristics of PV power generation being dynamically affected by sunlight and weather. To address this, this embodiment sets up a continuous constant value anomaly detection mechanism. The system traverses the power time series and counts the duration of adjacent equal values. If the duration exceeds a preset threshold (e.g., 60 minutes, corresponding to 4 15-minute sampling points), it is determined to be an "abnormal continuous constant value".
[0030] To address this deficiency, two approaches are employed: First, a daily cycle model reconstruction is used: extracting the typical output curves of the substation or feeder during the same historical period (same season, same weather type) and normalizing and scaling them according to the current day's maximum power, then replacing the data from the abnormal period. Second, a meteorological-driven reconstruction is used: if the meteorological data (such as irradiance) shows significant changes (e.g., irradiance fluctuations caused by cloud movement), but the power remains constant, then based on the approximate linear relationship between irradiance and power, the power value of the abnormal period is re-estimated using the conversion coefficient of the normal period (e.g., power generation efficiency per unit irradiance).
[0031] 4. Determine whether the power generation time exceeds the reasonable sunshine period. If it does, mark it as invalid and fill in the missing data.
[0032] Photovoltaic power generation relies on solar radiation, and its output should theoretically be close to zero at night, before sunrise, or for a very short period after sunset. However, due to factors such as meter clock deviation, data acquisition system malfunctions, or communication delays, the actual collected distributed photovoltaic power data may show anomalies such as non-zero power generation recorded during periods without sunshine (e.g., late at night). Therefore, it is necessary to determine whether the power generation time exceeds the reasonable sunshine period.
[0033] The filling here is divided into two cases: first, if it is clearly nighttime, then it is directly set to zero; second, if it is in the transition zone before sunrise / after sunset, then the power decay / rise curve of the same period in history is referenced for interpolation filling.
[0034] 5. Determine if there are any abnormal power generation trends. If so, fill in the data based on the power change patterns of normal days in the same historical period.
[0035] Photovoltaic power generation output typically exhibits a smooth daily variation curve. Under conditions of no significant meteorological disturbances (such as sudden cloud cover), its power change rate should be within a reasonable range. This embodiment analyzes the power change slope or first-order difference sequence for the current time period using a sliding window. If a continuous monotonically increasing or decreasing trend is detected, and the rate of change exceeds a preset threshold (e.g., a power change rate greater than 20% / hour for three consecutive 15-minute intervals), it is determined to be an abnormal power generation trend (such anomalies are usually caused by sensor drift, inverter control failure, or data transmission errors). For this anomaly, the system retrieves records from the historical database of "normal days" with similar seasons and weather types (e.g., sunny / cloudy) without anomaly markers. It extracts the typical power change pattern (i.e., normalized output profile) for the corresponding time period, scales it according to the current day's maximum power, and uses it to replace the data for the abnormal time period, thus achieving physically consistent and temporally reasonable intelligent filling.
[0036] 6. Determine whether there are any upper or lower boundary anomalies in the meteorological data. If so, truncate or fill with nearby valid values based on the preset threshold range of the meteorological data.
[0037] Meteorological data (such as irradiance, temperature, and humidity) have clear physical limits and reasonable ranges for different regions. This scheme sets preset threshold ranges for each type of meteorological data (e.g., irradiance of 0–1360 W / m²). 2 The ambient temperature is -40°C to +60°C, and the relative humidity is 0% to 100%. This range is determined based on the solar constant, typical climate extremes in China, and sensor range. The system iterates through the input meteorological time series. If the value at a certain moment exceeds the preset threshold range of the corresponding element (e.g., recorded at 1500 W / m³),... 2 If the irradiance is above a certain threshold or the temperature is below 70°C, it is considered an upper or lower bound anomaly. For such anomalies, the system preferentially uses truncation processing, that is, it forcibly corrects the value to the nearest boundary value (e.g., 1500 W / m²). 2 → 1360 W / m 2 If the duration of the anomaly is long or the boundary truncation may introduce bias, then the strategy of filling with the nearest effective value is adopted. That is, linear interpolation or sliding window mean is used to use the effective meteorological observations that are adjacent to each other to restore the continuity and rationality of the data and ensure the reliability of subsequent feature construction and prediction model input.
[0038] Step S3. Based on the training data, align the historical power generation data with the meteorological data at the corresponding time according to the timestamp, and extract the historical power generation lag features, the meteorological features of the same period, the time context features and the derived features based on the preset time window, and combine them to form an initial feature set. Then, use the feature construction algorithm to select the feature subset of the initial feature set to form the optimal feature subset.
[0039] The preset time window includes several past time steps, i.e. discrete time units divided with a fixed time granularity (e.g., 15 minutes); the historical power generation lag characteristics include power generation values at several time points before the current moment; the concurrent meteorological characteristics include forecast irradiance, temperature, and humidity at the current moment and several future time points; the time context characteristics include one or more of the following: hour identifier, weekday type, seasonal factor, or whether it is a holiday; the derived characteristics include the first-order difference or rate of change of power generation, clear sky index (the ratio of measured irradiance to theoretical maximum irradiance), and the moving average or trend slope of meteorological data.
[0040] The feature subset selection includes the following process: S31. Assign scores or weights to each feature in the initial feature set using the feature construction algorithm; S32. Sort the scores from high to low, and select the top k features in sequence to form a candidate subset. The value of k can be multiple, such as k=5, 10, 15, etc., to form multiple candidate subsets. S33. Input multiple candidate subsets into the validation model for cross-validation and calculate the corresponding prediction error index. The prediction error index includes one or more of the following: root mean square error, mean absolute percentage error, or continuous ranking probability score. The validation model is implemented using existing machine learning algorithm frameworks, such as decision trees, random forests, or gradient boosting tree models. These algorithms can directly call the standard implementations in open-source machine learning libraries (such as scikit-learn and XGBoost) without additional development. S34. Select the candidate subset that minimizes the prediction error index as the optimal feature subset.
[0041] The system integrates historical power generation lag terms, concurrent and forecast meteorological elements, time context (such as hours, seasons, and holidays), and derived features (such as power change rate, clear sky index, meteorological moving average, and trend slope) in a spatiotemporal alignment to form an initial feature set. A feature subset selection mechanism based on cross-validation is then used to automatically select the optimal feature combination with the smallest prediction error from multiple candidate subsets, balancing model accuracy and complexity and significantly improving feature representation capabilities.
[0042] Step S4. Using the optimal feature subset, train the prediction model using machine learning algorithms (such as decision tree, neural network, support vector machine, gradient boosting tree or one or more).
[0043] Step S5. Based on the trained prediction model, perform multi-timescale rolling prediction and output the predicted point values and corresponding confidence level probability interval prediction results within the preset future time period.
[0044] The multiple time scales here include ultra-short-term (i.e., the next 4 hours), short-term (i.e., the next 3 days), and medium-term (i.e., the next 7 days). The probability interval prediction results include at least the upper and lower bounds of the prediction intervals corresponding to three confidence levels: 85%, 90%, and 95%. This probabilistic result provides dispatchers with a basis for risk quantification and can be flexibly applied to different business scenarios such as reserve capacity configuration, maintenance plan optimization, and safety margin assessment.
[0045] Step S6. Publish the probability interval prediction results to the outside world for use by the scheduling and management system.
[0046] This embodiment also proposes a distributed photovoltaic power generation prediction system, including: The data acquisition module is used to acquire historical power generation data of distributed photovoltaic power and meteorological data from the electricity consumption information collection system and the meteorological system, respectively. The data quality control module is used to identify and repair anomalies in the acquired data and generate training data. The feature engineering module is used to build and optimize the feature set for training the prediction model, forming the optimal feature subset. The model training module is used to train the prediction model based on the optimal feature subset; The prediction execution module is used to call the trained prediction model to perform rolling predictions at multiple time scales and output point prediction values and probability interval prediction results at different confidence levels. The results publishing module is used to provide services to external parties with the probability interval prediction results.
[0047] The above embodiments are only used to explain the concept of the present invention, and are not intended to limit the protection of the present invention. Any non-substantial modifications made to the present invention using this concept should fall within the protection scope of the present invention.
Claims
1. A method for predicting distributed photovoltaic power generation, characterized in that, Includes the following processes: S1. Obtain historical power generation data of distributed photovoltaics and meteorological data of corresponding spatiotemporal locations, and perform time alignment on the acquired data to form a multidimensional time-series dataset; S2. Perform anomaly identification and repair processing on the multidimensional time-series dataset to generate training data; S3. Based on the training data, the historical power generation data and the meteorological data at the corresponding time are aligned by timestamp, and the historical power generation lag features, the meteorological features of the same period and the time context features are extracted based on the preset time window, and combined to form an initial feature set. Then, the feature construction algorithm is used to select the feature subset of the initial feature set to form the optimal feature subset. S4. Using the aforementioned optimal feature subset, a machine learning algorithm is employed to train a prediction model; S5. Based on the trained prediction model, perform multi-timescale rolling prediction and output the predicted point values and corresponding confidence level probability interval prediction results within the preset future time period. S6. Publish the probability interval prediction results to the outside world for use by the scheduling and management system.
2. The distributed photovoltaic power generation prediction method as described in claim 1, characterized in that, In S1, the meteorological data includes gridded irradiance, temperature, and humidity data, with a data time interval of 15 minutes; the historical power generation data comes from the electricity consumption information collection system and is aggregated by transformer area or feeder.
3. The distributed photovoltaic power generation prediction method as described in claim 1, characterized in that, The anomaly identification and repair process in S2 includes one or more of the following: Determine if there are missing values; if so, fill them using interpolation or the mean of nearby time periods. Alternatively, it can determine whether there are any abnormal mutations, and if so, make corrections based on sliding window trend detection; Alternatively, it can determine whether there are continuous constant value anomalies. If they exist and the duration exceeds the threshold, it can be reconstructed based on the daily cycle pattern or meteorological changes. Alternatively, it can determine whether the power generation time exceeds the reasonable sunshine period; if it does, it should be marked as invalid and filled in. Or determine whether there is an abnormal power generation trend; if so, fill in the data based on the power change pattern of normal days in the same historical period. Alternatively, it can determine whether there are upper or lower boundary anomalies in the meteorological data. If so, it can truncate or fill the data with nearby valid values based on the preset threshold range of the meteorological data.
4. The distributed photovoltaic power generation prediction method as described in claim 1, characterized in that, In S3, the preset time window includes several past time steps; The historical power generation lag characteristics include the power generation values at several time points prior to the current moment; The meteorological characteristics mentioned above include the forecast irradiance, temperature, and humidity at the current moment and several future time points; The time context features include one or more of the following: hour identifier, weekday type, seasonal factor, or whether it is a holiday.
5. The distributed photovoltaic power generation prediction method as described in claim 4, characterized in that, The initial feature set also includes derived features, which include one or more of the following: First-order difference or rate of change of power generation; Or the clear sky index, which is the ratio of measured irradiance to theoretical maximum irradiance; Or the moving average or trend slope of meteorological data.
6. The distributed photovoltaic power generation prediction method as described in claim 1, characterized in that, The feature construction algorithm in S3 includes at least one of recursive feature elimination, feature importance ranking based on a tree model, or principal component analysis.
7. The distributed photovoltaic power generation prediction method as described in claim 1, characterized in that, In step S3, the feature subset selection includes the following process: S31. Assign scores or weights to each feature in the initial feature set using the feature construction algorithm; S32. Sort the scores from high to low, and select the top k features in sequence to form a candidate subset, where k can take multiple values to form multiple candidate subsets; S33. Input multiple candidate subsets into the validation model for cross-validation and calculate the corresponding prediction error index. The prediction error index includes one or more of the following: root mean square error, mean absolute percentage error, or continuous ranking probability score. S34. Select the candidate subset that minimizes the prediction error index as the optimal feature subset.
8. The distributed photovoltaic power generation prediction method as described in claim 1, characterized in that, In S5, the multiple time scales include: Ultra-short term, meaning the next 4 hours; Short term, meaning the next 3 days; The medium term refers to the next 7 days.
9. The distributed photovoltaic power generation prediction method as described in claim 1, characterized in that, The probability interval prediction results include at least the upper and lower bounds of the prediction intervals corresponding to the three confidence levels of 85%, 90%, and 95%.
10. A distributed photovoltaic power generation prediction system based on the method of any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to acquire historical power generation data of distributed photovoltaic power and meteorological data from the electricity consumption information collection system and the meteorological system, respectively. The data quality control module is used to identify and repair anomalies in the acquired data and generate training data. The feature engineering module is used to build and optimize the feature set for training the prediction model, forming the optimal feature subset. The model training module is used to train the prediction model based on the optimal feature subset; The prediction execution module is used to call the trained prediction model to perform rolling predictions at multiple time scales and output point prediction values and probability interval prediction results at different confidence levels. The results publishing module is used to provide services to external parties with the probability interval prediction results.