Photovoltaic output correction prediction method and system based on multi-source heterogeneous data

Abstract

The invention discloses a photovoltaic output correction prediction method and system based on multi-source heterogeneous data, and belongs to the technical field of photovoltaic output prediction.The method comprises the following steps that after the multi-source heterogeneous data are preprocessed and integrated, a load feature data set is extracted; constructing a photovoltaic output prediction model; obtaining control group data in the database to obtain a control group prediction result; analyzing the sensitivity of the photovoltaic output prediction model to each control group; extracting a feature type corresponding to the sensitivity not lower than the sensitivity threshold; and analyzing the precision value of the feature data subsequently acquired by the feature type. According to the method, the change sensitivity of each feature type in the photovoltaic output prediction model is analyzed, the feature type with high sensitivity is subjected to precision detection, and whether the feature type can be put into the model or not is determined according to a precision detection result. According to the method, the overall accuracy and robustness of the photovoltaic output prediction model can be improved, and the problem that the prediction precision of the model is reduced due to interference of low-quality features is solved.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic power output prediction technology, and in particular relates to a photovoltaic power output correction prediction method and system based on multi-source heterogeneous data. Background Technology

[0002] With the rapid development and large-scale application of photovoltaic power generation technology, distributed photovoltaics has become an important component of modern power systems. Against the backdrop of global energy structure transformation, photovoltaic installed capacity has been increasing year by year. However, photovoltaic output has significant intermittency and volatility. Especially after distributed photovoltaics are connected to the grid, their randomness and uncertainty place higher demands on the stability and security of the power grid.

[0003] Existing photovoltaic (PV) power output forecasting technologies mainly rely on statistical modeling based on meteorological and historical power output data, and machine learning model training. For example, CN117993739A discloses a short-term PV power output forecasting method based on data denoising and point correction, employing different forecasting models and utilizing the control variable method for short-term PV power output forecasting. CN117689087B discloses a PV power station power output forecasting method and system for solar eclipse events, using a corrected power output model to predict the dynamic changes in PV power station power output during the passage of a solar eclipse.

[0004] However, when using existing technologies to predict photovoltaic power, due to data heterogeneity, complex operating scenarios, and real-time fluctuations, the model input features may be complex and subject to interference from low-quality features, leading to decreased model prediction accuracy and insufficient reliability. Summary of the Invention

[0005] The purpose of this invention is to provide a photovoltaic power output correction prediction method and system based on multi-source heterogeneous data, which solves the problem of decreased prediction accuracy and insufficient reliability of existing models due to interference from low-quality features, and improves the overall prediction accuracy and robustness of photovoltaic power output prediction models.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A photovoltaic power output correction prediction method based on multi-source heterogeneous data includes the following steps: After acquiring multi-source heterogeneous data and preprocessing and integrating the multi-source heterogeneous data, the load feature dataset is extracted. A photovoltaic power output prediction model was constructed based on the load characteristic dataset. Retrieve the control group data of the control variables set according to the feature type in the database, input the control group data into the photovoltaic output prediction model, and obtain the prediction results of the control group; Based on the data and prediction results of each control group, the sensitivity of the photovoltaic power output prediction model to each control group was analyzed. Sensitivity is the magnitude of change in the photovoltaic power output prediction result when the characteristic data changes. The sensitivity is compared with the sensitivity threshold in the database, and the feature types corresponding to the sensitivity that is not lower than the sensitivity threshold are extracted. For each extracted feature type, a precision check is set. The precision value of the feature data subsequently acquired for that feature type is analyzed. If the precision value is higher than the precision threshold set in the database, it is allowed to be used in the photovoltaic power output prediction model. Otherwise, it is not allowed to be used in the photovoltaic power output prediction model, and the feature data of that feature type is notified to be optimized.

[0007] Preferably, the process of preprocessing and integrating multi-source heterogeneous data is as follows: Acquire multi-source heterogeneous data, including distributed photovoltaic data, low-voltage distribution area data at the end of the distribution network, energy storage data, and environmental monitoring data; Multi-source heterogeneous data is time-aligned according to timestamps and spatially aligned according to spatial location. Data completion and cleaning of multi-source heterogeneous data are performed, and the completed and cleaned multi-source heterogeneous data is then standardized to obtain preprocessed multi-source heterogeneous data.

[0008] Preferably, the method for extracting the load feature dataset is as follows: Based on the preprocessed multi-source heterogeneous data, each data is used as an independent feature vector. The encoder in the autoencoder compresses the input features into a low-dimensional space to extract latent features. The latent features are restored from the low-dimensional space to the original high-dimensional space using the decoder in the autoencoder to obtain the load feature dataset.

[0009] Preferably, the steps for constructing the photovoltaic output prediction model are as follows: The load characteristic dataset includes photovoltaic load characteristics and photovoltaic output characteristics; The photovoltaic load features in the preprocessed load feature dataset are used as input features, and the photovoltaic output features in the load feature dataset are used as prediction targets. The load feature dataset is sorted according to time series, and then the dataset is divided into training and test sets. The training set is input into the multinomial regression model for training to obtain a preliminary photovoltaic power output prediction model. The test set is input into the preliminary photovoltaic power output prediction model for testing to obtain a photovoltaic power output prediction model.

[0010] Preferably, the control group data set according to the feature type is as follows: For the feature type of the load feature dataset, multiple sets of data samples are constructed based on the control variable method: Each set of samples allows only the value of one feature type to fluctuate, while the values ​​of the other feature types remain unchanged, so as to isolate the influence of a single feature on the photovoltaic output prediction results.

[0011] Preferably, the process of analyzing the sensitivity of the photovoltaic power output prediction model to each control group is as follows: Obtain the input data of each feature type of multi-source heterogeneous data in the photovoltaic power output prediction model, as well as the output data corresponding to each input data. Record the input data of the same feature type and its corresponding output data as the sub-data group of that feature type. Obtain the feature types and their total number, the sub-data groups for each feature type and their total number, and then standardize them; The feature weights of each feature type in the photovoltaic power output prediction model are obtained through an objective weighting method. Based on the standardized feature types and their total number, and the sub-data sets of each feature type and their total number, the sensitivity of the photovoltaic power output prediction model to the control group corresponding to each feature type is calculated using the sensitivity calculation formula.

[0012] The preferred formula for sensitivity calculation is: ; In the formula, Let represent the sensitivity of the control group corresponding to the i-th feature type, where i is the feature type number. I represents the total number of feature type numbers, and k represents the sub-data group number for each feature type. K is the total number of sub-data group numbers for each feature type. This is the input data for the k-th sub-data group in the i-th feature type. This is the output data of the k-th sub-data group in the i-th feature type. This is the input data for the (k-1)th sub-data group in the i-th feature type. This is the output data for the (k-1)th sub-data group in the i-th feature type. Let be the weight of the i-th feature type in relation to sensitivity.

[0013] Preferably, the steps for accuracy detection are as follows: The accuracy of power measurement sensors, environmental monitoring sensors, and energy storage system sensors were obtained from the sensor manuals and then standardized. The feature weights of the accuracy values ​​of power measurement sensors, environmental monitoring sensors, and energy storage system sensors are obtained through an objective weighting method. The accuracy value of the control group corresponding to each feature type is calculated using the accuracy value calculation formula.

[0014] Preferably, the formula for calculating the accuracy value is: ; In the formula, JDZ is the accuracy value, and DLCL is the accuracy of the power measurement sensor. DLCL represents the weighting factor of JDZ, and HJJC represents the accuracy of the environmental monitoring sensor. HJJC is the weighting factor of JDZ, and CNXT represents the sensor accuracy of the energy storage system. This is the weighting factor of CNXT on JDZ.

[0015] A photovoltaic output correction and prediction system based on multi-source heterogeneous data, used to implement the above method, includes: The data preprocessing module is used to acquire multi-source heterogeneous data, and after preprocessing and integrating the multi-source heterogeneous data, extract the load feature dataset. The prediction model building module is used to build a photovoltaic output prediction model based on the load characteristic dataset. The control group data acquisition module is used to acquire control group data of control variables set according to feature type in the database, input the control group data into the photovoltaic output prediction model, and obtain the control group prediction results. The sensitivity analysis module is used to analyze the sensitivity of the photovoltaic power output prediction model to each control group based on the data and prediction results of each control group. Sensitivity refers to the magnitude of change in the photovoltaic power output prediction result when the feature data changes. The sensitivity filtering module is used to compare the sensitivity with the sensitivity threshold in the database and extract the feature type corresponding to the sensitivity that is not lower than the sensitivity threshold. The accuracy detection and optimization notification module is used to set accuracy detection for the extracted feature types, analyze the accuracy value of the feature data subsequently obtained for the feature type, and allow the feature to be used in the photovoltaic power output prediction model when the accuracy value is higher than the accuracy threshold set in the database; otherwise, it is not allowed to be used in the photovoltaic power output prediction model, and the module will notify the optimization of the feature data for the feature type.

[0016] The beneficial effects of this invention are as follows: This invention constructs a control group of data and conducts sensitivity analysis based on the control group data, thereby accurately identifying the key features affecting the results of the photovoltaic power output prediction model. This enables the priority analysis of highly sensitive features and the exclusion of low-sensitivity features, optimizing the complexity and computational efficiency of the prediction model and solving the problem of decreased model prediction accuracy caused by interference from low-quality features.

[0017] This invention uses an accuracy detection and optimization notification module to dynamically evaluate and filter the accuracy of data of highly sensitive feature types, thereby ensuring that the data input into the prediction model has high accuracy characteristics. This improves the reliability of the model prediction results and controls errors, avoiding large errors in the model prediction results due to inconsistent feature data quality.

[0018] This invention improves the accuracy and reliability of photovoltaic power output prediction models and reduces model iteration and optimization costs by determining whether high-sensitivity feature types meet accuracy requirements and notifying optimization accordingly. Attached Figure Description

[0019] Figure 1 This is a schematic flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the modules of the system of the present invention. Detailed Implementation

[0020] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0021] Example 1: This example addresses the problem of decreased model prediction accuracy caused by interference from low-quality features. The overall approach is as follows: By acquiring multi-source heterogeneous data and preprocessing and integrating it, a load characteristic dataset is extracted; a photovoltaic power output prediction model is constructed; control group data is obtained from the database, and the prediction results of the control group are obtained; the sensitivity of the photovoltaic power output prediction model to each control group is analyzed; feature types corresponding to sensitivity not lower than the sensitivity threshold are extracted; and the accuracy value of subsequent feature data acquisition for this feature type is analyzed, thereby improving the overall accuracy and robustness of the photovoltaic power output prediction model.

[0022] like Figure 1 As shown, a photovoltaic output correction prediction method based on multi-source heterogeneous data includes the following steps: After acquiring multi-source heterogeneous data and preprocessing and integrating the multi-source heterogeneous data, the load feature dataset is extracted. A photovoltaic power output prediction model was constructed based on the load characteristic dataset. Retrieve the control group data of the control variables set according to the feature type in the database, input the control group data into the photovoltaic output prediction model, and obtain the prediction results of the control group; Based on the data and prediction results of each control group, the sensitivity of the photovoltaic power output prediction model to each control group was analyzed. Sensitivity is the magnitude of change in the photovoltaic power output prediction result when the characteristic data changes. The sensitivity is compared with the sensitivity threshold in the database, and the feature types corresponding to the sensitivity that is not lower than the sensitivity threshold are extracted. For each extracted feature type, a precision check is set. The precision value of the feature data subsequently acquired for that feature type is analyzed. If the precision value is higher than the precision threshold set in the database, it is allowed to be used in the photovoltaic power output prediction model. Otherwise, it is not allowed to be used in the photovoltaic power output prediction model, and the feature data of that feature type is notified to be optimized.

[0023] Acquire multi-source heterogeneous data, including distributed photovoltaic data, low-voltage distribution area data at the end of the distribution network, energy storage data, and environmental monitoring data. Clean and standardize the data to remove noise and missing values. Extract load feature datasets from the preprocessed data. These features include the power output of the photovoltaic system, meteorological data (such as temperature, humidity, and light intensity), and load changes. Use the extracted features as input data for the photovoltaic output prediction model.

[0024] A predictive model is established using a regression model based on machine learning. Control variable data for the control group is obtained based on the feature type. The control group data is then input into the photovoltaic output prediction model to generate the prediction results for the control group. By analyzing the control group data and its corresponding prediction results, the sensitivity of the model to each feature can be determined. Based on the sensitivity of each feature, it can be identified which feature changes will significantly affect the prediction results of photovoltaic output.

[0025] Feature types with sensitivity values ​​higher than or equal to a sensitivity threshold are extracted, and their accuracy is tested. If the accuracy of the feature data meets the set accuracy threshold, it is allowed to be used in the photovoltaic output prediction model. High-accuracy feature data improves the model's sensitivity to changes in the external environment, ensuring stable performance under different photovoltaic operating scenarios. If the accuracy does not meet the requirements, it is not used, and relevant personnel are notified to optimize the feature data, eliminating low-quality data to prevent it from introducing interference into the model. This ensures the reliability of the model's prediction results from the source, avoids the negative impact of low-quality features on the model, and continuously improves data quality through optimized feedback mechanisms. In addition, rejecting feature data below the accuracy threshold from being used in the model also avoids processing invalid data, reduces the number of model training iterations, improves system operating efficiency, and reduces operating costs.

[0026] Sensitivity analysis allows for the dynamic assessment of the impact of each feature on the prediction model, optimizing the model's predictive capabilities. By setting accuracy thresholds, it ensures that only high-quality data can be incorporated into the prediction model, which helps reduce data noise and improve the reliability and stability of predictions. For features that significantly impact prediction results, accuracy detection and feedback mechanisms can be used to optimize feature data until the required accuracy is met before deployment, thereby enhancing the system's long-term performance.

[0027] The specific steps for preprocessing and integrating multi-source heterogeneous data include: Acquire multi-source heterogeneous data, including distributed photovoltaic data, low-voltage distribution area data at the end of the distribution network, energy storage data, and environmental monitoring data; align the multi-source heterogeneous data by time according to timestamps and by spatial location; perform data completion and cleaning on the multi-source heterogeneous data, and then standardize the completed and cleaned multi-source heterogeneous data to obtain preprocessed multi-source heterogeneous data.

[0028] Time alignment refers to aligning the time steps of multi-source heterogeneous data, such as unifying them to minute or hour levels; spatial alignment refers to matching the physical regions belonging to different data sources based on geocoding or power grid topology relationships in multi-source heterogeneous data, such as matching the spatial location information of photovoltaic sites and energy storage devices in distribution areas. Data completion can be accomplished using linear interpolation or Lagrange interpolation methods; data cleaning can use algorithms such as the Isolation Forest to remove outliers.

[0029] Time alignment of multi-source heterogeneous data ensures that different data sources can be compared at the same time granularity, avoiding data bias caused by time inconsistency, and facilitating real-time monitoring, modeling, and prediction in subsequent analysis.

[0030] Spatial alignment of multi-source heterogeneous data enables data to be compared and analyzed according to actual geographical locations, avoiding analytical errors caused by spatial location differences. This helps to understand the different data characteristics and patterns in different spatial locations and improves the prediction accuracy of the model.

[0031] The method for extracting the load feature dataset is as follows: based on the preprocessed multi-source heterogeneous data, each data is used as an independent feature vector. The encoder in the autoencoder compresses the input features into a low-dimensional space to extract the latent features. The decoder in the autoencoder restores the latent features from the low-dimensional space to the original high-dimensional space to obtain the load feature dataset.

[0032] First, based on the preprocessed multi-source heterogeneous data, each data type is input as an independent feature vector into the autoencoder model. In the autoencoder, the encoder part compresses the input high-dimensional feature vector into a low-dimensional latent feature space through a series of nonlinear mappings. The latent features represent the core features and structural information of the input data. Subsequently, the decoder part performs an inverse nonlinear mapping on the latent features to restore the low-dimensional features to the high-dimensional space, recover the structural information of the original data, and generate a load feature dataset. This dataset integrates the features and latent characteristics of the multi-source data and is suitable for subsequent analysis and modeling.

[0033] The steps for constructing the photovoltaic power output prediction model are as follows: The load feature dataset includes photovoltaic load features and photovoltaic power output features; the photovoltaic load features in the preprocessed load feature dataset are used as input features, and the photovoltaic power output features in the load feature dataset are used as prediction targets; the load feature dataset is sorted according to time series and divided into training and testing sets; the training set is input into a multinomial regression model for training to obtain a preliminary photovoltaic power output prediction model; the testing set is input into the preliminary photovoltaic power output prediction model for testing to obtain the final photovoltaic power output prediction model.

[0034] The load feature dataset is sorted according to time series. The sorted load feature dataset is then divided according to a set ratio. Typically, a larger proportion of the data is allocated as the training set for model training, and the remainder is allocated as the test set for model validation. The training set is then input into a multinomial regression model, which is used to fit the training set to find the multinomial relationship between the input features and the prediction target, thus obtaining a preliminary photovoltaic power output prediction model. Subsequently, the test set is input into the preliminary photovoltaic power output prediction model to verify its performance and predictive ability. Based on the test results, the model is evaluated and optimized as necessary to obtain the final photovoltaic power output prediction model.

[0035] The control group data set according to the feature type refers to: for the feature type of the load feature dataset, multiple sets of data samples are constructed based on the control variable method: each set of samples is allowed to have fluctuations in the value of only one feature type, while the values ​​of the other feature types remain unchanged, so as to isolate the influence of a single feature on the photovoltaic output prediction results.

[0036] By grouping the load characteristic dataset, each group contains complete data for all characteristic types; within each group, one characteristic type is selected as a variable, while the remaining characteristic types are kept at their original values; the selected characteristic types are adjusted multiple times according to the variable's value range or reference data stored in the database to form multiple control variable control group data; the generated control group data are recorded; through this step, independent analysis for a single characteristic type can be achieved, thereby improving the sensitivity evaluation accuracy of the photovoltaic power output prediction model.

[0037] The steps for analyzing the sensitivity of the photovoltaic power output prediction model to each control group are as follows: First, obtain the input data and corresponding output data for each feature type in the photovoltaic power output prediction model; denote the input data and corresponding output data of the same feature type as sub-data groups of that feature type. Second, obtain the feature types and their total number, the sub-data groups of each feature type and their total number, and standardize them. Third, obtain the feature weights of each feature type in the photovoltaic power output prediction model using an objective weighting method. Fourth, calculate the sensitivity of the photovoltaic power output prediction model to the control group corresponding to each feature type based on the standardized feature types and their total number, and the sub-data groups of each feature type and their total number, using the sensitivity calculation formula. The sensitivity calculation formula is: ; In the formula, Let represent the sensitivity of the control group corresponding to the i-th feature type, where i is the feature type number. I represents the total number of feature type numbers, and k represents the sub-data group number for each feature type. K is the total number of sub-data group numbers for each feature type. This is the input data for the k-th sub-data group in the i-th feature type. This is the output data of the k-th sub-data group in the i-th feature type. This is the input data for the (k-1)th sub-data group in the i-th feature type. This is the output data for the (k-1)th sub-data group in the i-th feature type. Let be the weight of the i-th feature type in relation to sensitivity.

[0038] Feature types include, but are not limited to: Distributed photovoltaic data: The operating data of the photovoltaic system itself (such as the output power of photovoltaic modules, the operating status of inverters, etc.); Data on low-voltage distribution areas at the end of the distribution network: Operational data on the low-voltage side of the distribution network (such as distribution area load, voltage / current fluctuations, line losses, etc.). Energy storage data: Data related to energy storage systems (such as energy storage charging and discharging power, remaining power, and operating status of energy storage devices). Environmental monitoring data: External environmental data that affects photovoltaic output (such as light intensity, temperature, humidity, wind speed, etc.).

[0039] An exemplary calculation process is as follows: I=1, K=2, =7, =9, =3, =6, =1, therefore... =0.559, The higher the value, the higher the sensitivity of the photovoltaic power output prediction model to the control group corresponding to this feature type; conversely, the lower the value, the lower the sensitivity.

[0040] By comparing sensitivity with a set sensitivity threshold, feature data that significantly impacts photovoltaic power output prediction results is selected, ensuring the accuracy of key variables input into the model meets the standards. Furthermore, by selecting feature types that are at least below the sensitivity threshold, unnecessary interference from low-sensitivity features can be eliminated, reducing the computational load of irrelevant or weakly correlated data and improving model efficiency and resource utilization.

[0041] The specific steps for accuracy testing are as follows: First, obtain the accuracy of the power measurement sensor, environmental monitoring sensor, and energy storage system sensor from the sensor manuals, and perform standardization processing. Second, obtain the feature weights of the power measurement sensor accuracy, environmental monitoring sensor accuracy, and energy storage system sensor accuracy with respect to the accuracy value using an objective weighting method. Third, calculate the accuracy value of the corresponding control group for each feature type using the accuracy value calculation formula. The accuracy value calculation formula is as follows: ; In the formula, JDZ is the accuracy value, and DLCL is the accuracy of the power measurement sensor. DLCL represents the weighting factor of JDZ, and HJJC represents the accuracy of the environmental monitoring sensor. HJJC is the weighting factor of JDZ, and CNXT represents the sensor accuracy of the energy storage system. This is the weighting factor of CNXT on JDZ.

[0042] An example calculation process is as follows: DLCL is 0.8. =0.4, HJJC=0.7, =0.3, CNXT=0.9, =0.3, resulting in JDZ=0.624. The higher the JDZ value, the higher the accuracy value of the feature corresponding to the control group, and vice versa.

[0043] By setting precision checks on the extracted feature types, the reliability of key feature data is ensured, effectively reducing the risk of prediction bias caused by low-quality data, improving the output accuracy and stability of the prediction model, and preventing low-precision data from entering the model to avoid introducing noise or erroneous information, thereby ensuring the overall performance of the prediction model. The notification mechanism can provide timely feedback on optimization needs, helping to continuously improve the data collection and processing flow, forming a virtuous cycle.

[0044] Example 2: When constructing a photovoltaic power output prediction model, a dynamic feature selection mechanism based on reinforcement learning is introduced to adaptively optimize the feature set and improve the model's prediction performance. Specifically, this includes: Dynamic Feature Selection Mechanism: In the process of constructing the photovoltaic power output prediction model, a reinforcement learning framework is introduced. Through the interaction between the intelligent agent and the environment (the photovoltaic power output prediction task), the most influential features on the prediction results are dynamically selected. The intelligent agent takes actions (selecting or removing features) based on the current state (model prediction error, feature importance, etc.) and adjusts its strategy based on the reward signals (improvement or decrease in prediction accuracy) from the environment.

[0045] Reinforcement learning framework design: A Markov Decision Process (MDP) is used to model the feature selection problem. The state space includes the current model's prediction error, the set of selected features, and the set of unselected features. The action space represents selecting or removing a feature. The reward function is designed to improve the model's prediction accuracy, i.e., the reduction in prediction error after selecting or removing a feature. By training the reinforcement learning model, the intelligent agent can learn the optimal feature selection strategy.

[0046] Synergy between feature selection and model optimization: In each iteration, the intelligent agent selects a subset of features based on the current strategy, then retrains the photovoltaic power output prediction model using this subset and calculates the new prediction accuracy. A positive reward is given if the prediction accuracy improves; otherwise, a negative reward is given. Through multiple iterations, the intelligent agent gradually optimizes the feature selection strategy, ultimately selecting the feature set that has the greatest impact on photovoltaic power output prediction, thereby significantly improving the model's prediction performance.

[0047] By introducing a dynamic feature selection mechanism based on reinforcement learning, adaptive optimization of feature selection is achieved. This method not only automatically adjusts the feature set according to dynamic changes in the data, but also evaluates the effectiveness of features in real time during model training, thus avoiding the bias that may be introduced by manually pre-setting the feature set. Through dynamic optimization of the feature set, the model can better adapt to complex, multi-source, heterogeneous data environments, significantly improving prediction accuracy and robustness.

[0048] Example 3: As Figure 2 As shown, a photovoltaic output correction and prediction system based on multi-source heterogeneous data is used to implement the method in Example 1 or Example 2, including: The data preprocessing module is used to acquire multi-source heterogeneous data, and after preprocessing and integrating the multi-source heterogeneous data, extract the load feature dataset. The prediction model building module is used to build a photovoltaic output prediction model based on the load characteristic dataset. The control group data acquisition module is used to acquire control group data of control variables set according to feature type in the database, input the control group data into the photovoltaic output prediction model, and obtain the control group prediction results. The sensitivity analysis module is used to analyze the sensitivity of the photovoltaic power output prediction model to each control group based on the data and prediction results of each control group. Sensitivity refers to the magnitude of change in the photovoltaic power output prediction result when the feature data changes. The sensitivity filtering module is used to compare the sensitivity with the sensitivity threshold in the database and extract the feature type corresponding to the sensitivity that is not lower than the sensitivity threshold. The accuracy detection and optimization notification module is used to set accuracy detection for the extracted feature types, analyze the accuracy value of the feature data subsequently obtained for the feature type, and allow the feature to be used in the photovoltaic power output prediction model when the accuracy value is higher than the accuracy threshold set in the database; otherwise, it is not allowed to be used in the photovoltaic power output prediction model, and the module will notify the optimization of the feature data for the feature type.

[0049] Methods for optimizing feature data of this feature type include: Upgrade the acquisition equipment or sensors to improve the accuracy of hardware measurements. For example, replace the acquisition equipment with high-precision equipment such as high-resolution radiometers, temperature sensors, or current sensors, or regularly calibrate existing equipment to reduce errors caused by equipment aging or environmental influences.

[0050] For highly sensitive features, more data processing and acquisition resources are dynamically allocated. For example, a higher sampling frequency is set for highly sensitive features during the data acquisition process to improve the accuracy of key feature data.

[0051] By combining multiple data sources to perform cross-validation and comparative analysis on data with the same characteristics, the most reliable data can be selected. For example, reference measurement data can be used to adjust the deviation of the collected data, or standard weather station data can be used to correct the field sensor data, thereby improving data consistency and reliability.

[0052] Statistical or machine learning algorithms are used to label and correct outliers, missing values, or noise in the original data. For example, filtering algorithms (such as mean filtering and median filtering) are applied to smooth data curves, or distribution fitting models are used to impute missing values, thereby reducing data bias and ensuring the authenticity of the input data. Through these optimization methods, the accuracy of feature-type data can be further improved, ensuring the performance and stability of the photovoltaic power output prediction model.

Claims

1. A photovoltaic power output correction and prediction method based on multi-source heterogeneous data, characterized in that, Includes the following steps: After acquiring multi-source heterogeneous data and preprocessing and integrating the multi-source heterogeneous data, the load feature dataset is extracted. A photovoltaic power output prediction model was constructed based on the load characteristic dataset. Retrieve the control group data of the control variables set according to the feature type in the database, input the control group data into the photovoltaic output prediction model, and obtain the prediction results of the control group; Based on the data and prediction results of each control group, the sensitivity of the photovoltaic power output prediction model to each control group was analyzed. Sensitivity is the magnitude of change in the photovoltaic power output prediction result when the characteristic data changes. The sensitivity is compared with the sensitivity threshold in the database, and the feature types corresponding to the sensitivity that is not lower than the sensitivity threshold are extracted. For each extracted feature type, a precision check is set. The precision value of the feature data subsequently acquired for that feature type is analyzed. If the precision value is higher than the precision threshold set in the database, it is allowed to be used in the photovoltaic power output prediction model. Otherwise, it is not allowed to be used in the photovoltaic power output prediction model, and the feature data of that feature type is notified to be optimized.

2. The photovoltaic output correction and prediction method based on multi-source heterogeneous data as described in claim 1, characterized in that, The process of preprocessing and integrating multi-source heterogeneous data is as follows: Acquire multi-source heterogeneous data, including distributed photovoltaic data, low-voltage distribution area data at the end of the distribution network, energy storage data, and environmental monitoring data; Multi-source heterogeneous data is time-aligned according to timestamps and spatially aligned according to spatial location. Data completion and cleaning of multi-source heterogeneous data are performed, and the completed and cleaned multi-source heterogeneous data is then standardized to obtain preprocessed multi-source heterogeneous data.

3. The photovoltaic output correction and prediction method based on multi-source heterogeneous data as described in claim 1, characterized in that, The method for extracting the load feature dataset is as follows: Based on the preprocessed multi-source heterogeneous data, each data is used as an independent feature vector. The encoder in the autoencoder compresses the input features into a low-dimensional space to extract latent features. The latent features are restored from the low-dimensional space to the original high-dimensional space using the decoder in the autoencoder to obtain the load feature dataset.

4. The photovoltaic output correction and prediction method based on multi-source heterogeneous data as described in claim 1, characterized in that, The steps for constructing a photovoltaic power output prediction model are as follows: The load characteristic dataset includes photovoltaic load characteristics and photovoltaic output characteristics; The photovoltaic load features in the preprocessed load feature dataset are used as input features, and the photovoltaic output features in the load feature dataset are used as prediction targets. The load feature dataset is sorted according to time series, and then the dataset is divided into training and test sets. The training set is input into the multinomial regression model for training to obtain a preliminary photovoltaic power output prediction model. The test set is input into the preliminary photovoltaic power output prediction model for testing to obtain a photovoltaic power output prediction model.

5. The photovoltaic output correction and prediction method based on multi-source heterogeneous data as described in claim 1, characterized in that, The control group data set according to the feature type is as follows: For the feature type of the load feature dataset, multiple sets of data samples are constructed based on the control variable method: Each set of samples allows only the value of one feature type to fluctuate, while the values ​​of the other feature types remain unchanged, so as to isolate the influence of a single feature on the photovoltaic output prediction results.

6. The photovoltaic output correction and prediction method based on multi-source heterogeneous data as described in claim 1, characterized in that, The process of analyzing the sensitivity of the photovoltaic power output prediction model to each control group is as follows: Obtain the input data of each feature type of multi-source heterogeneous data in the photovoltaic power output prediction model, as well as the output data corresponding to each input data. Record the input data of the same feature type and its corresponding output data as the sub-data group of that feature type. Obtain the feature types and their total number, the sub-data groups for each feature type and their total number, and then standardize them; The feature weights of each feature type in the photovoltaic power output prediction model are obtained through an objective weighting method. Based on the standardized feature types and their total number, and the sub-data sets of each feature type and their total number, the sensitivity of the photovoltaic power output prediction model to the control group corresponding to each feature type is calculated using the sensitivity calculation formula.

7. The photovoltaic output correction and prediction method based on multi-source heterogeneous data as described in claim 6, characterized in that, The formula for calculating sensitivity is: ； In the formula, Let represent the sensitivity of the control group corresponding to the i-th feature type, where i is the feature type number. I represents the total number of feature type numbers, and k represents the sub-data group number for each feature type. K is the total number of sub-data group numbers for each feature type. This is the input data for the k-th sub-data group in the i-th feature type. This is the output data of the k-th sub-data group in the i-th feature type. This is the input data for the (k-1)th sub-data group in the i-th feature type. This is the output data for the (k-1)th sub-data group in the i-th feature type. Let be the weight of the i-th feature type in relation to sensitivity.

8. The photovoltaic output correction and prediction method based on multi-source heterogeneous data as described in claim 2, characterized in that, The steps for accuracy testing are as follows: The accuracy of power measurement sensors, environmental monitoring sensors, and energy storage system sensors were obtained from the sensor manuals and then standardized. The feature weights of the accuracy values ​​of power measurement sensors, environmental monitoring sensors, and energy storage system sensors are obtained through an objective weighting method. The accuracy value of the control group corresponding to each feature type is calculated using the accuracy value calculation formula.

9. The photovoltaic output correction and prediction method based on multi-source heterogeneous data as described in claim 8, characterized in that, The formula for calculating the accuracy value is: ； In the formula, JDZ is the accuracy value, and DLCL is the accuracy of the power measurement sensor. DLCL represents the weighting factor of JDZ, and HJJC represents the accuracy of the environmental monitoring sensor. HJJC is the weighting factor of JDZ, and CNXT represents the sensor accuracy of the energy storage system. This is the weighting factor of CNXT on JDZ.

10. A photovoltaic output correction and prediction system based on multi-source heterogeneous data, used to implement the method described in any one of claims 1-9, characterized in that, include: The data preprocessing module is used to acquire multi-source heterogeneous data, and after preprocessing and integrating the multi-source heterogeneous data, extract the load feature dataset. The prediction model building module is used to build a photovoltaic output prediction model based on the load characteristic dataset. The control group data acquisition module is used to acquire control group data of control variables set according to feature type in the database, input the control group data into the photovoltaic output prediction model, and obtain the control group prediction results. The sensitivity analysis module is used to analyze the sensitivity of the photovoltaic power output prediction model to each control group based on the data and prediction results of each control group. Sensitivity refers to the magnitude of change in the photovoltaic power output prediction result when the feature data changes. The sensitivity filtering module is used to compare the sensitivity with the sensitivity threshold in the database and extract the feature type corresponding to the sensitivity that is not lower than the sensitivity threshold. The accuracy detection and optimization notification module is used to set accuracy detection for the extracted feature types, analyze the accuracy value of the feature data subsequently obtained for the feature type, and allow the feature to be used in the photovoltaic power output prediction model when the accuracy value is higher than the accuracy threshold set in the database; otherwise, it is not allowed to be used in the photovoltaic power output prediction model, and the module will notify the optimization of the feature data for the feature type.

Images