A mountain photovoltaic PR intelligent analysis method based on big data fusion
By using a heterogeneous tree structure based on big data fusion and online asynchronous update technology, the tree splitting rules are dynamically adjusted, solving the problem of data non-uniformity in mountain photovoltaic power stations and achieving efficient and accurate performance ratio prediction and evaluation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- YUNNAN DIANENG SMART ENERGY CO LTD
- Filing Date
- 2026-04-20
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies struggle to accurately reflect the operational status of different regions in mountainous photovoltaic power plants due to the non-uniformity and complex spatiotemporal distribution of data. Traditional models also have limited generalization capabilities and low training efficiency.
We employ a big data fusion approach, using heterogeneous tree structures and online asynchronous update technology to dynamically adjust tree splitting rules and optimize parameters. We then use the CatBoost and FTRL algorithms to train the gradient boosting decision tree model.
It improves the accuracy and training efficiency of performance ratios for mountain photovoltaic power stations, ensures high adaptability and accuracy of the model in complex environments, and avoids waste of computing resources and instability during the training process.
Smart Images
Figure CN122432669A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent analysis technology for photovoltaic power generation, and in particular to an intelligent analysis method for mountain photovoltaic power generation based on big data fusion. Background Technology
[0002] With the continuous expansion of the construction scale of mountain photovoltaic power stations, the evaluation of power generation efficiency and the analysis of operational status under complex terrain conditions have gradually become an important research direction in the photovoltaic industry. Performance ratio (PR), as a key indicator for measuring the difference between the actual and theoretical power generation capacity of a photovoltaic power station, is widely used in power station operation evaluation, fault diagnosis, and operation and maintenance decision-making. Existing technologies typically collect meteorological data, component operating parameters, and historical power generation data to perform statistical analysis of PR or make predictions and evaluations based on traditional machine learning models. Mountain photovoltaic power stations are significantly affected by factors such as slope aspect, altitude changes, local climate differences, and irradiance fluctuations. The data exhibits strong non-uniformity and complex spatiotemporal distribution characteristics, making it difficult for a single model to accurately reflect the operational status of different regions.
[0003] In existing technologies, most PR analysis methods still employ a unified data modeling approach, mixing data from different time periods, weather conditions, and geographical regions for training. This lack of a fine-grained segmentation mechanism to address data differences can easily lead to insufficient learning of local features. Traditional gradient boosting decision trees or synchronously trained models typically rely on fixed splitting rules for tree construction, making it difficult to adapt to variations in multi-source data distribution and limiting model generalization capabilities. Existing distributed training methods often employ synchronous update strategies, which are easily affected by performance differences among computing nodes in multi-data source scenarios, resulting in decreased training efficiency.
[0004] Therefore, how to provide a smart analysis method for mountain photovoltaic PR based on big data fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] One objective of this invention is to propose an intelligent performance ratio (PR) analysis method for mountainous photovoltaic (PV) power plants based on big data fusion. This invention fully utilizes heterogeneous tree structures and online asynchronous update technology, combined with dynamically adjusted tree splitting rules and efficient parameter optimization. This method can accurately evaluate the performance ratio of mountainous PV power plants and has the advantages of high adaptability, high accuracy, and high training efficiency.
[0006] A mountain photovoltaic PR intelligent analysis method based on big data fusion according to an embodiment of the present invention includes the following steps: Obtain the original training dataset, and perform data partitioning on the original training dataset to generate several data subsets; For each subset of data, an asynchronous training process is started, and each asynchronous training process is executed independently and in parallel; In each asynchronous training process, an initial gradient boosting decision tree model is constructed based on the CatBoost algorithm framework, and the statistical feature information of the current data subset is obtained. The tree splitting rule combination corresponding to the current data subset is dynamically determined based on the statistical feature information. A heterogeneous tree structure is constructed by combining tree splitting rules. The heterogeneous tree structure is a decision tree generated by selecting the corresponding splitting rule from the combination of tree splitting rules at each layer of the decision tree according to the preset layer feature selection strategy to split the nodes. Gradient calculation is performed on the current data subset based on the heterogeneous tree structure to obtain local gradient information. The FTRL algorithm is then used to update the local model parameters of the current asynchronous training process online based on the local gradient information, generating the updated local model parameters. After reaching the preset global synchronization period, the independent training of all asynchronous training processes is paused. The local model parameters generated by the asynchronous training processes are weighted and averaged to generate global model parameters, which are then used to replace all local model parameters. Repeat the training until the preset training convergence condition is met, and output the final target gradient boosting decision tree model.
[0007] Optionally, generating several data subsets specifically includes: Analyze the original training dataset and extract the sampling time point and weather type label for each training sample. The original training dataset is divided into several time slice data blocks according to the time order of the sampling time points; For each time slice data block, training samples with the same weather type label are grouped into the same weather group based on the weather type label; For each weather group, based on the geographical zoning information of the mountain photovoltaic power station, the training samples located in the same mountain slope area and the same altitude range are further divided into the corresponding geographical subdivision data blocks; Each time slice data block, each weather group, and each geographic subdivision data block is treated as an independent subset of data.
[0008] Optionally, the asynchronous training process may further include a lifecycle management step: Monitor the running status of each asynchronous training process; When the continuous unresponsive duration of an asynchronous training process exceeds the preset heartbeat timeout threshold, the current asynchronous training process is determined to be invalid, the invalid asynchronous training process is terminated, and the occupied computing resources are released. Reread the data subset corresponding to the failed process from the distributed storage, restart a new asynchronous training process to replace the failed asynchronous training process, and load the latest global model parameters into the new asynchronous training process; Record the cumulative runtime and completed training rounds of each asynchronous training process. When the cumulative runtime of an asynchronous training process exceeds the preset maximum runtime threshold, forcibly terminate the asynchronous training process and mark the data subset corresponding to the process as having completed training. When all asynchronous training processes corresponding to all data subsets are marked as completed, the final model parameter fusion is triggered during the global synchronization cycle.
[0009] Optionally, the dynamic determination of the tree splitting rule combination corresponding to the current data subset specifically includes: Perform statistical analysis on each training sample in the current data subset and calculate the statistical feature vector; Based on the statistical feature vector, the dispersion index of continuous features and the balance index of category features are calculated. When the dispersion index of continuous features is higher than the preset dispersion threshold, the tree splitting rule is added to the tree splitting rule combination based on the quantile interval segmentation rule; when the dispersion index of continuous features is lower than the dispersion threshold, the tree splitting rule is added to the tree splitting rule combination based on the mean neighborhood difference. The class distribution status is determined based on the class feature balance index. When the class distribution entropy value is higher than the preset balance threshold, the class splitting rule based on information gain is added to the tree splitting rule combination; when the class distribution entropy value is lower than the balance threshold, the class splitting rule based on the Gini index is added to the tree splitting rule combination. The tree splitting rule combination is validated and screened. The current data subset is input into the initial gradient boosting decision tree model to perform a pre-training calculation. The node loss reduction amount corresponding to different splitting rules is recorded. The splitting rules that satisfy the preset reduction ratio condition are retained to form the final tree splitting rule combination.
[0010] Optionally, the construction of a heterogeneous tree structure using tree splitting rules specifically includes: Using the current data subset as input, a root node is generated in the initial gradient boosting decision tree model, and the node splitting rule for the root node is determined from the final tree splitting rule combination according to the preset layer feature selection strategy and the node splitting is executed. After generating child nodes, perform statistical analysis on the sample distribution of the current layer nodes. When the node loss of the current layer nodes decreases to meet the preset preservation condition, continue to execute node splitting according to the node splitting rule adopted by the parent node. When the node loss of the current layer node does not meet the preset preservation condition, a new node splitting rule is determined from the final tree splitting rule combination based on the sample statistical characteristics of the current layer node and the node splitting is performed. When the final tree splitting rule combination contains more than one candidate splitting rule, the node splitting rule for the current layer node is determined according to the preset rule priority; when the final tree splitting rule combination contains only a single splitting rule, the node splitting is directly performed using the single splitting rule. Repeat the node splitting rule determination and node splitting process until the preset maximum tree depth is reached or the number of node samples meets the stop splitting condition, and generate a heterogeneous tree structure.
[0011] Optionally, the online asynchronous update of the local model parameters specifically includes: Input the current subset of data into the heterogeneous tree structure to perform forward prediction calculation, and generate local gradient information for the current training round based on the error between the prediction result and the corresponding training label; According to the sample partitioning path of each node in the heterogeneous tree structure, the local gradient information is allocated to the corresponding node to form a node gradient set; In the current asynchronous training process, the FTRL algorithm is invoked to generate parameter update amounts based on the cumulative gradient information of historical training rounds and the local gradient information of the current training round, and to perform online updates on the local model parameters. After the local model parameters are updated, the update results are directly applied to the node parameters corresponding to the current heterogeneous tree structure, and the updated local model parameters participate in the gradient calculation of subsequent training rounds. Each asynchronous training process independently executes the above update process without waiting for other training processes to complete their updates, generating updated local model parameters.
[0012] Optionally, the generation of global model parameters specifically includes: During the independent online update process of each asynchronous training process, the current training round is recorded. When all asynchronous training processes reach the preset global synchronization period, a synchronization control command is sent to each asynchronous training process to pause the independent training operation of each asynchronous training process. Collect the updated local model parameters corresponding to each asynchronous training process, and determine the parameter fusion weights based on the number of training samples corresponding to each data subset; The local model parameters of each asynchronous training process are weighted and averaged according to the fusion weights to generate the global model parameters; The generated global model parameters are distributed to each asynchronous training process, replacing the currently saved local model parameters in each asynchronous training process; After the local model parameters are replaced, the independent training of each asynchronous training process is restarted, so that the updated global model parameters can participate in the gradient calculation and online update of the next training cycle.
[0013] Optionally, the training convergence condition specifically includes: After each global synchronization cycle is completed, the node loss reduction ratio of the corresponding data subset of each asynchronous training process between the current training round and the previous training round is calculated. When the proportion of the number of data subsets with a node loss reduction ratio lower than the first preset threshold to the total number of data subsets is greater than or equal to the preset consistency ratio, it is determined that the consistency convergence condition is met. The difference between the global model parameters generated in two consecutive global synchronization cycles is calculated. When the parameter change between the two global model parameters is less than the second preset threshold, it is determined that the parameter stability convergence condition is met. When both the consistency convergence condition and the parameter stability convergence condition are satisfied, the training of each asynchronous training process is stopped, and the final target gradient boosting decision tree model is output.
[0014] The beneficial effects of this invention are: (1) By dividing the data into subsets based on multiple dimensions such as weather type, geographical location, and time window, each data subset can adopt different tree splitting rules according to its own characteristics, avoiding the neglect of complex environmental features by traditional models and improving the accuracy and robustness of the model in mountain photovoltaic power stations.
[0015] (2) Through the asynchronous training process and the online asynchronous update mechanism of FTRL, the present invention can process multiple data subsets in parallel without waiting for other processes to complete, reducing the waste of computing resources and effectively improving the training efficiency of large-scale datasets, ensuring rapid response and real-time adjustment in practical applications.
[0016] (3) By introducing a global synchronization cycle and multiple convergence determination mechanisms, the parameter update process of the heterogeneous tree structure in multiple asynchronous processes is ensured to be stable and reliable. By jointly determining the convergence of multiple subsets and the stable convergence of global model parameters, premature stopping and convergence instability during training are avoided, ensuring the high accuracy and high reliability of the final output model. Attached Figure Description
[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 The flowchart shows a mountain photovoltaic PR intelligent analysis method based on big data fusion proposed in this invention. Figure 2 This diagram illustrates the initiation and parallel execution of the asynchronous training process of a mountain photovoltaic PR intelligent analysis method based on big data fusion proposed in this invention. Figure 3 This is a schematic diagram of the FTRL algorithm update for a mountain photovoltaic PR intelligent analysis method based on big data fusion proposed in this invention. Detailed Implementation
[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0019] refer to Figures 1-3 A smart analysis method for the PR of mountain photovoltaic power generation based on big data fusion includes the following steps: Obtain the original training dataset, and perform data partitioning on the original training dataset to generate several data subsets; For each subset of data, an asynchronous training process is started, and each asynchronous training process is executed independently and in parallel; In each asynchronous training process, an initial gradient boosting decision tree model is constructed based on the CatBoost algorithm framework, and the statistical feature information of the current data subset is obtained. The tree splitting rule combination corresponding to the current data subset is dynamically determined based on the statistical feature information. A heterogeneous tree structure is constructed by combining tree splitting rules. The heterogeneous tree structure is a decision tree generated by selecting the corresponding splitting rule from the combination of tree splitting rules at each layer of the decision tree according to the preset layer feature selection strategy to split the nodes. Gradient calculation is performed on the current data subset based on the heterogeneous tree structure to obtain local gradient information. The FTRL algorithm is then used to update the local model parameters of the current asynchronous training process online based on the local gradient information, generating the updated local model parameters. After reaching the preset global synchronization period, the independent training of all asynchronous training processes is paused. The local model parameters generated by the asynchronous training processes are weighted and averaged to generate global model parameters, which are then used to replace all local model parameters. Repeat the training until the preset training convergence condition is met, and output the final target gradient boosting decision tree model.
[0020] In this embodiment, generating several data subsets specifically includes: The original training dataset was analyzed to extract the sampling time point and weather type label for each training sample. Weather conditions were categorized as sunny, cloudy, and overcast based on the proportion of total cloud cover to the sky. Specifically, a total cloud cover of less than 20% was labeled as sunny, a total cloud cover between 20% and 80% was labeled as cloudy, and a total cloud cover greater than 80% with no precipitation was labeled as overcast. When precipitation was detected, it was labeled as rain or snow regardless of cloud cover level. According to the time sequence of the sampling time points, the original training dataset is divided into several time slice data blocks; the time window length is set to 1 hour, and the 24 hours of the day are divided into 24 consecutive and non-overlapping time windows starting from 0:00 every day. For each time slice data block, training samples with the same weather type label are grouped into the same weather group based on the weather type label; For each weather group, based on the geographical zoning information of the mountain photovoltaic power station, training samples located in the same mountain slope area and the same altitude range are further divided into corresponding geographical subdivision data blocks; the altitude is discretized into intervals of 100 meters, and the slope angle is divided into four main directions: north slope, east slope, south slope, and west slope; when the same training sample meets multiple division conditions at the same time, the sample will be copied and assigned to multiple data subsets. Each time slice data block, each weather group, and each geographic subdivision data block is treated as an independent data subset. A corresponding irradiance level label and temperature range label are added to each data subset. The average irradiance value corresponding to this data subset is extracted from historical irradiance data and divided into low, medium, and high irradiance levels according to preset thresholds. The average ambient temperature corresponding to this data subset is extracted from historical meteorological data and divided into temperature range labels in 5°C increments.
[0021] In this embodiment, the asynchronous training process also includes a lifecycle management step: Monitor the running status of each asynchronous training process; When the continuous unresponsive duration of an asynchronous training process exceeds the preset heartbeat timeout threshold, the current asynchronous training process is determined to be invalid, the invalid asynchronous training process is terminated, and the occupied computing resources are released; in this embodiment, the heartbeat timeout threshold is set to 30 seconds. Reread the data subset corresponding to the failed process from the distributed storage, restart a new asynchronous training process to replace the failed asynchronous training process, and load the latest global model parameters into the new asynchronous training process; The cumulative runtime and completed training rounds of each asynchronous training process are recorded. When the cumulative runtime of an asynchronous training process exceeds the preset maximum runtime threshold, the asynchronous training process is forcibly terminated and the data subset corresponding to the process is marked as having completed training. In this embodiment, the maximum runtime threshold is set to 24 hours. The specific values of the heartbeat timeout threshold and the maximum runtime can be adjusted according to the data scale and computing resources of the photovoltaic power station, with adjustment ranges of 10 seconds to 60 seconds and 12 hours to 48 hours, respectively. When all asynchronous training processes corresponding to all data subsets are marked as completed, the final model parameter fusion is triggered during the global synchronization cycle.
[0022] In this embodiment, dynamically determining the tree splitting rule combination corresponding to the current data subset specifically includes: Statistical analysis is performed on each training sample in the current data subset to calculate the mean, standard deviation, skewness, and quantile distribution parameters of continuous features. At the same time, the number of categories, the proportion of category frequency, and the category distribution entropy value of categorical features are statistically analyzed to form a statistical feature vector of the data subset. Based on statistical feature vectors, the dispersion index of continuous features and the balance index of categorical features are calculated. When the dispersion index of continuous features is higher than a preset dispersion threshold, a tree splitting rule based on quantile intervals is added to the tree splitting rule combination; when the dispersion index of continuous features is lower than the dispersion threshold, a splitting rule based on the mean neighborhood difference is added to the tree splitting rule combination. For the continuous feature dispersion index, the standard deviation σ, interquartile range (IQR), and skewness are calculated for each continuous feature. The ratio of the standard deviation to the mean of the continuous feature, the IQR, and the feature value range are then compared. The absolute value weighted sum of ewness is used as an index of the dispersion of continuous features; the dispersion threshold is set to 0.65; the quantile interval segmentation rule refers to generating candidate segmentation points according to the quantiles of continuous features, dividing the samples into four intervals according to the 25%, 50%, and 75% quantiles, and selecting the quantile that causes the node loss function to decrease the most as the node splitting position; the mean neighborhood difference segmentation rule refers to constructing neighborhood intervals within the mean ± kσ range of the continuous feature mean, and calculating the difference between the mean of the target variable of samples inside and outside the neighborhood; The category distribution status is determined based on the category feature balance index. When the category distribution entropy value is higher than the preset balance threshold, the category splitting rule based on information gain is added to the tree splitting rule combination; when the category distribution entropy value is lower than the balance threshold, the category splitting rule based on Gini index is added to the tree splitting rule combination. The category feature balance index is defined as the category distribution entropy value. In this embodiment, the balance threshold is taken as 70% of the maximum entropy value of the category. The category splitting rule based on information gain selects the category splitting method with the largest information gain as the node splitting method by calculating the decrease in information entropy of the target variable before and after node splitting. The category splitting rule based on Gini index selects the category splitting method with the smallest Gini index as the node splitting method by calculating the Gini impurity after splitting. The tree splitting rule combinations are validated and filtered. The current data subset is input into the initial gradient boosting decision tree model to perform a pre-training calculation. The reduction in node loss corresponding to different splitting rules is recorded. The splitting rules whose node loss reduction satisfies the preset reduction ratio condition are retained to form the final tree splitting rule combination. In this embodiment, the preset reduction ratio threshold is set to 0.08.
[0023] In this embodiment, the construction of a heterogeneous tree structure using tree splitting rules specifically includes: Using the current data subset as input, a root node is generated in the initial gradient boosting decision tree model, and the node splitting rule for the root node is determined from the final tree splitting rule combination according to the preset layer feature selection strategy and the node splitting is executed. After generating child nodes, statistical analysis is performed on the sample distribution of the current layer nodes. When the node loss of the current layer node decreases to meet the preset preservation condition, the node splitting rule adopted by the parent node is maintained and node splitting continues. The preset preservation condition is that the node loss decrease ratio is greater than or equal to 0.05. When the node loss of the current layer node does not meet the preset preservation condition, a new node splitting rule is determined from the final tree splitting rule combination based on the sample statistical characteristics of the current layer node and the node splitting is performed. When the final tree splitting rule combination contains more than one candidate splitting rule, the node splitting rule for the current layer node is determined according to the preset rule priority; when the final tree splitting rule combination contains only a single splitting rule, the single splitting rule is directly used to perform node splitting; in this embodiment, the preset rule priority order is: continuous feature splitting rules take precedence over category feature splitting rules; among continuous feature splitting rules, the splitting rule with a larger reduction in node loss takes precedence; among category feature splitting rules, the splitting rule with a larger increase in node purity takes precedence. Repeat the node splitting rule determination and node splitting process until the preset maximum tree depth is reached or the number of node samples meets the stopping condition, generating a heterogeneous tree structure. The preset maximum tree depth is 8 levels, and the stopping condition is that node splitting stops when the number of node samples is less than 50.
[0024] In this embodiment, the online asynchronous update of local model parameters specifically includes: The current data subset is input into the heterogeneous tree structure to perform forward prediction calculation. Based on the error between the prediction result and the corresponding training label, the local gradient information of the current training round is generated; the error is calculated using the squared loss function. According to the sample partitioning path of each node in the heterogeneous tree structure, the local gradient information is allocated to the corresponding node to form a node gradient set; In the current asynchronous training process, the FTRL algorithm is invoked to generate parameter update amounts based on the cumulative gradient information of historical training rounds and the local gradient information of the current training round, and online updates are performed on the local model parameters; updates stop when the loss decreases by less than 0.01 for two consecutive rounds of local updates. After the local model parameters are updated, the update results are directly applied to the node parameters corresponding to the current heterogeneous tree structure, and the updated local model parameters participate in the gradient calculation of subsequent training rounds. Each asynchronous training process independently executes the above update process without waiting for other training processes to complete their updates, generating updated local model parameters.
[0025] In this embodiment, generating global model parameters specifically includes: During the independent online update process of each asynchronous training process, the current training round is recorded. When all asynchronous training processes reach the preset global synchronization period, a synchronization control command is sent to each asynchronous training process to pause the independent training operation of each asynchronous training process. The global synchronization period is measured in training rounds. Global synchronization is performed once every 5 to 20 rounds of local training. In this embodiment, 10 rounds are used. Collect the updated local model parameters corresponding to each asynchronous training process, and determine the parameter fusion weight based on the number of training samples corresponding to each data subset; the fusion weight is determined according to the proportion of the number of training samples in each data subset to the total number of training samples. The local model parameters of each asynchronous training process are weighted and averaged according to the fusion weights to generate the global model parameters; The generated global model parameters are distributed to each asynchronous training process, replacing the currently saved local model parameters in each asynchronous training process; After the local model parameters are replaced, the independent training of each asynchronous training process is restarted, so that the updated global model parameters can participate in the gradient calculation and online update of the next training cycle.
[0026] In this embodiment, the training convergence conditions specifically include: After each global synchronization cycle is completed, the percentage decrease in node loss for each asynchronous training process's corresponding data subset between the current training round and the previous training round is calculated. When the proportion of data subsets with a node loss decrease ratio below a first preset threshold to the total number of data subsets is greater than or equal to a preset consistency ratio, the consistency convergence condition is determined to be met. The first preset threshold is 0.02; the preset consistency ratio is 0.80, meaning that when at least 80% of the data subsets satisfy the condition that the node loss decrease ratio is below the first preset threshold, the consistency convergence condition is determined to be met. The difference between the global model parameters generated in two consecutive global synchronization cycles is calculated. When the change in parameters between the two global model parameters is less than a second preset threshold, it is determined that the parameter stability convergence condition is met. The second preset threshold is used to determine whether the global model parameters generated in two consecutive global synchronization cycles are stable. The second preset threshold is 5×10-4. The change in parameters is calculated using the L2 norm difference between the global model parameter vectors of two consecutive cycles. When both the consistency convergence condition and the parameter stability convergence condition are satisfied, the training of each asynchronous training process is stopped, and the final target gradient boosting decision tree model is output.
[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to the performance ratio (PR) intelligent analysis of a mountain photovoltaic power generation system. The aim was to address the problems of low accuracy and poor adaptability of traditional performance ratio (PR) assessment methods for mountain photovoltaic power plants due to complex geographical environments and meteorological conditions. Mountain photovoltaic power plants are affected by factors such as terrain and climate, resulting in data characteristics exhibiting strong non-uniformity and complex spatiotemporal distribution. Existing technologies struggle to accurately analyze and model this variable data.
[0028] This invention introduces a multi-feature analysis method based on big data fusion, combined with the CatBoost algorithm and the FTRL online update mechanism, enabling accurate prediction and analysis of performance (PR) under complex geographical conditions. By dividing data into different time periods, weather types, and geographical regions, and dynamically selecting the most suitable splitting rule using a heterogeneous tree structure, the model's high adaptability to different data features is ensured. Each data subset is executed independently during training, and the global model parameters are finally generated through global synchronization and weighted averaging, significantly improving training efficiency and the model's prediction accuracy.
[0029] In an experimental application at a mountain photovoltaic power station, we analyzed the station's historical data. This power station is distributed across multiple geographical areas, facing varying weather types and altitudes. In the experiment, we integrated historical meteorological data, real-time data from the photovoltaic modules, and power generation data, and performed data preprocessing, subset partitioning, and asynchronous training according to the method described in this invention.
[0030] In the experiment, we first cleaned the raw data through data preprocessing, removing missing and outlier values, and then divided the data according to weather conditions, time windows, and geographical regions. The dataset was divided into four main subsets: sunny, cloudy, overcast, and rainy / snowy, with each subset containing corresponding temperature, humidity, solar radiation intensity, wind speed, and component output data.
[0031] During asynchronous training, we constructed an initial gradient boosting decision tree model based on the CatBoost framework and dynamically selected the tree splitting rule according to the characteristics of each data subset. For example, on the sunny data subset, we chose an information gain-based splitting rule, while on the cloudy data subset, we adopted a Gini index-based splitting rule. During asynchronous training on each data subset, the FTRL algorithm was used to update the local model parameters online, optimizing model performance by continuously adjusting the learning rate and regularization parameters.
[0032] The table below shows a comparison of the model's performance, prediction error, and training time under different weather conditions before and after applying the method of this invention. The data table illustrates the difference in PR prediction accuracy between traditional methods and the method of this invention, as well as the improvement in training efficiency.
[0033] Table 1: Comparison of PR forecast performance for weather conditions
[0034] As shown in Table 1, the PR prediction error of the method of this invention is lower than that of the traditional method under all weather conditions, especially under sunny and cloudy weather, the prediction error is reduced by 7.2% and 8.1%, respectively. Furthermore, the training time using the method of this invention is also shortened compared to the traditional method; under sunny conditions, the training time is reduced from 45 minutes to 30 minutes, and the training efficiency is improved by 33%. This result indicates that the intelligent PR analysis method for mountain photovoltaic systems based on big data fusion can not only improve prediction accuracy but also enhance training efficiency, making it particularly suitable for real-time performance evaluation and fault diagnosis of large-scale photovoltaic power plants.
[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A smart analysis method for mountain photovoltaic PR based on big data fusion, characterized in that, Includes the following steps: Obtain the original training dataset, and perform data partitioning on the original training dataset to generate several data subsets; For each subset of data, an asynchronous training process is started, and each asynchronous training process is executed independently and in parallel; In each asynchronous training process, an initial gradient boosting decision tree model is constructed based on the CatBoost algorithm framework, and the statistical feature information of the current data subset is obtained. The tree splitting rule combination corresponding to the current data subset is dynamically determined based on the statistical feature information. A heterogeneous tree structure is constructed by combining tree splitting rules. The heterogeneous tree structure is a decision tree generated by selecting the corresponding splitting rule from the combination of tree splitting rules at each layer of the decision tree according to the preset layer feature selection strategy to split the nodes. Gradient calculation is performed on the current data subset based on the heterogeneous tree structure to obtain local gradient information. The FTRL algorithm is then used to update the local model parameters of the current asynchronous training process online based on the local gradient information, generating the updated local model parameters. After reaching the preset global synchronization period, the independent training of all asynchronous training processes is paused. The local model parameters generated by the asynchronous training processes are weighted and averaged to generate global model parameters, which are then used to replace all local model parameters. Repeat the training until the preset training convergence condition is met, and output the final target gradient boosting decision tree model.
2. The intelligent analysis method for mountain photovoltaic PR based on big data fusion according to claim 1, characterized in that, The generation of several data subsets specifically includes: Analyze the original training dataset and extract the sampling time point and weather type label for each training sample. The original training dataset is divided into several time slice data blocks according to the time order of the sampling time points; For each time slice data block, training samples with the same weather type label are grouped into the same weather group based on the weather type label; For each weather group, based on the geographical zoning information of the mountain photovoltaic power station, the training samples located in the same mountain slope area and the same altitude range are further divided into the corresponding geographical subdivision data blocks; Each time slice data block, each weather group, and each geographic subdivision data block is treated as an independent subset of data.
3. The intelligent analysis method for mountain photovoltaic PR based on big data fusion according to claim 2, characterized in that, The asynchronous training process also includes lifecycle management steps: Monitor the running status of each asynchronous training process; When the continuous unresponsive duration of an asynchronous training process exceeds the preset heartbeat timeout threshold, the current asynchronous training process is determined to be invalid, the invalid asynchronous training process is terminated, and the occupied computing resources are released. Reread the data subset corresponding to the failed process from the distributed storage, restart a new asynchronous training process to replace the failed asynchronous training process, and load the latest global model parameters into the new asynchronous training process; Record the cumulative runtime and completed training rounds of each asynchronous training process. When the cumulative runtime of an asynchronous training process exceeds the preset maximum runtime threshold, forcibly terminate the asynchronous training process and mark the data subset corresponding to the process as having completed training. When all asynchronous training processes corresponding to all data subsets are marked as completed, the final model parameter fusion is triggered during the global synchronization cycle.
4. The intelligent analysis method for mountain photovoltaic PR based on big data fusion according to claim 3, characterized in that, The dynamic determination of the tree splitting rule combination corresponding to the current data subset specifically includes: Perform statistical analysis on each training sample in the current data subset and calculate the statistical feature vector; Based on the statistical feature vector, the dispersion index of continuous features and the balance index of category features are calculated. When the dispersion index of continuous features is higher than the preset dispersion threshold, the tree splitting rule is added to the tree splitting rule combination based on the quantile interval segmentation rule; when the dispersion index of continuous features is lower than the dispersion threshold, the tree splitting rule is added to the tree splitting rule combination based on the mean neighborhood difference. The class distribution status is determined based on the class feature balance index. When the class distribution entropy value is higher than the preset balance threshold, the class splitting rule based on information gain is added to the tree splitting rule combination; when the class distribution entropy value is lower than the balance threshold, the class splitting rule based on the Gini index is added to the tree splitting rule combination. The tree splitting rule combination is validated and screened. The current data subset is input into the initial gradient boosting decision tree model to perform a pre-training calculation. The node loss reduction amount corresponding to different splitting rules is recorded. The splitting rules that satisfy the preset reduction ratio condition are retained to form the final tree splitting rule combination.
5. The intelligent analysis method for mountain photovoltaic PR based on big data fusion according to claim 4, characterized in that, The specific methods for constructing heterogeneous tree structures using tree splitting rules include: Using the current data subset as input, a root node is generated in the initial gradient boosting decision tree model, and the node splitting rule for the root node is determined from the final tree splitting rule combination according to the preset layer feature selection strategy and the node splitting is executed. After generating child nodes, perform statistical analysis on the sample distribution of the current layer nodes. When the node loss of the current layer nodes decreases to meet the preset preservation condition, continue to execute node splitting according to the node splitting rule adopted by the parent node. When the node loss of the current layer node does not meet the preset preservation condition, a new node splitting rule is determined from the final tree splitting rule combination based on the sample statistical characteristics of the current layer node and the node splitting is performed. When the final tree splitting rule combination contains more than one candidate splitting rule, the node splitting rule for the current layer node is determined according to the preset rule priority; when the final tree splitting rule combination contains only a single splitting rule, the node splitting is directly performed using the single splitting rule. Repeat the node splitting rule determination and node splitting process until the preset maximum tree depth is reached or the number of node samples meets the stop splitting condition, and generate a heterogeneous tree structure.
6. The intelligent analysis method for mountain photovoltaic PR based on big data fusion according to claim 5, characterized in that, The online asynchronous update of the local model parameters specifically includes: Input the current subset of data into the heterogeneous tree structure to perform forward prediction calculation, and generate local gradient information for the current training round based on the error between the prediction result and the corresponding training label; According to the sample partitioning path of each node in the heterogeneous tree structure, the local gradient information is allocated to the corresponding node to form a node gradient set; In the current asynchronous training process, the FTRL algorithm is invoked to generate parameter update amounts based on the cumulative gradient information of historical training rounds and the local gradient information of the current training round, and to perform online updates on the local model parameters. After the local model parameters are updated, the update results are directly applied to the node parameters corresponding to the current heterogeneous tree structure, and the updated local model parameters participate in the gradient calculation of subsequent training rounds. Each asynchronous training process independently executes the above update process without waiting for other training processes to complete their updates, generating updated local model parameters.
7. The intelligent analysis method for mountain photovoltaic PR based on big data fusion according to claim 6, characterized in that, The generated global model parameters specifically include: During the independent online update process of each asynchronous training process, the current training round is recorded. When all asynchronous training processes reach the preset global synchronization period, a synchronization control command is sent to each asynchronous training process to pause the independent training operation of each asynchronous training process. Collect the updated local model parameters corresponding to each asynchronous training process, and determine the parameter fusion weights based on the number of training samples corresponding to each data subset; The local model parameters of each asynchronous training process are weighted and averaged according to the fusion weights to generate the global model parameters; The generated global model parameters are distributed to each asynchronous training process, replacing the currently saved local model parameters in each asynchronous training process; After the local model parameters are replaced, the independent training of each asynchronous training process is restarted, so that the updated global model parameters can participate in the gradient calculation and online update of the next training cycle.
8. The intelligent analysis method for mountain photovoltaic PR based on big data fusion according to claim 7, characterized in that, The training convergence conditions specifically include: After each global synchronization cycle is completed, the node loss reduction ratio of the corresponding data subset of each asynchronous training process between the current training round and the previous training round is calculated. When the proportion of the number of data subsets with a node loss reduction ratio lower than the first preset threshold to the total number of data subsets is greater than or equal to the preset consistency ratio, it is determined that the consistency convergence condition is met. The difference between the global model parameters generated in two consecutive global synchronization cycles is calculated. When the parameter change between the two global model parameters is less than the second preset threshold, it is determined that the parameter stability convergence condition is met. When both the consistency convergence condition and the parameter stability convergence condition are satisfied, the training of each asynchronous training process is stopped, and the final target gradient boosting decision tree model is output.