Distributed photovoltaic power anomaly identification and prediction method based on neural network
By combining neural network models, the problem of anomaly identification and prediction in distributed photovoltaic systems was solved, improving prediction accuracy and identification precision, and meeting real-time operation and maintenance requirements.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIAN THERMAL POWER RES INST CO LTD
- Filing Date
- 2025-07-23
- Publication Date
- 2026-05-19
AI Technical Summary
The diversity and concealment of anomalies in distributed photovoltaic systems make operation and maintenance difficult. Traditional methods are unable to capture nonlinear fluctuations in photovoltaic power and equipment aging, resulting in high missed detection rates, large prediction errors, and difficulty in meeting real-time requirements.
A neural network-based approach is adopted, which combines convolutional neural modules and recurrent neural modules to extract features from photovoltaic (PV) operation data and train models. This establishes a PV power anomaly identification and prediction model, including data mapping, curve construction, feature extraction, and model training, to achieve anomaly identification and prediction of PV systems.
It improves the accuracy of photovoltaic power prediction and anomaly identification, reduces the missed detection rate and prediction error, and enables real-time monitoring and timely handling of anomalies in photovoltaic systems.
Smart Images
Figure CN120892964B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of photovoltaic power anomaly identification and prediction, and in particular to a method and system for distributed photovoltaic power anomaly identification and prediction based on neural networks. Background Technology
[0002] Against the backdrop of the accelerated global energy transition to renewable energy, distributed photovoltaic (PV) power generation has become a crucial force in achieving the "dual carbon" goals due to its advantages such as local consumption, reduced transmission losses, and flexible grid connection. However, its inherent characteristics and complex application scenarios present significant technical challenges that severely restrict power generation efficiency and grid stability. From a physical perspective, the power output of distributed PV systems exhibits significant intermittency and randomness. Subtle changes in meteorological factors such as sunlight intensity, temperature, and cloud movement can lead to drastic nonlinear fluctuations in power output. For example, a sudden, short-duration heavy rainfall in the summer afternoon can cause the power output of PV modules to drop sharply to less than 10% of its peak value within minutes; while dew condensation in the early winter morning can also reduce power generation efficiency by decreasing the light transmittance of the modules. Furthermore, the coupling effect between sunlight intensity and temperature can trigger a "temperature compensation effect," further exacerbating the complexity of output characteristics.
[0003] At the system deployment level, the decentralized and small-scale nature of distributed photovoltaic (PV) systems further amplifies the difficulty of operation and maintenance. Individual power plants typically range in capacity from kilowatts to megawatts, widely distributed across user-side scenarios such as industrial parks, residential rooftops, and commercial buildings, resulting in a layout characterized by "numerous points, wide coverage, and diverse quantities." Statistics show that the average capacity of residential distributed PV projects in my country is less than 10kW, and most are located at the end of the grid. This not only leads to problems such as network latency and signal interference in data collection, but also makes it difficult for traditional operation and maintenance models based on centralized monitoring to meet real-time requirements. Simultaneously, issues such as inconsistent equipment models and fragmented communication protocols further exacerbate the complexity of data integration and collaborative management.
[0004] The diversity and concealment of anomalies have become core obstacles to the stable operation of distributed photovoltaic systems. The sources of these faults encompass three major dimensions: equipment, weather, and the power grid. At the equipment level, faults such as inverter over-temperature protection, photovoltaic module hot spot effects, and DC cable aging are progressive and insidious, initially manifesting only as a slight decrease in power, making them difficult to detect in a timely manner using threshold judgments. Regarding weather anomalies, sudden hailstorms, sandstorms, and localized shading caused by bird nesting can trigger module current mismatch, resulting in irreversible damage. Grid-side anomalies include voltage spikes / drops, harmonic pollution, and three-phase imbalances, which in severe cases may trigger anti-islanding protection, leading to unplanned outages.
[0005] Traditional technologies have significant limitations in addressing these challenges. Statistical methods, such as thresholding and time-series decomposition models, rely on fixed parameters and historical data patterns, making it difficult to capture nonlinear factors such as sudden changes in sunlight and equipment aging. Traditional machine learning methods, such as SVM and random forests, while possessing some nonlinear fitting capabilities, suffer from complex feature engineering and insufficient long-range dependency modeling when processing high-dimensional time-series data. For example, measured data from a distributed photovoltaic cluster in a certain region showed that the false negative rate was as high as 32% when using thresholding for anomaly detection, while the SVM model had an average absolute error exceeding 18% of the actual power when predicting power fluctuations under cloudy weather conditions.
[0006] To address these issues, there is an urgent need for a method and system for identifying and predicting anomalies in distributed photovoltaic power based on neural networks, in order to overcome the technical bottlenecks in the operation, maintenance, and prediction of distributed photovoltaic systems. Summary of the Invention
[0007] To address the aforementioned problems, this application proposes a method for identifying and predicting distributed photovoltaic power anomalies based on neural networks, comprising the following steps:
[0008] S1. Obtain past photovoltaic operation data, analyze the past operation data to obtain past environmental data and past power data, and match the past environmental data and past power data one by one to obtain the environment-power mapping group.
[0009] S2. Construct curves from the data in the environment-power mapping group to obtain a primary curve group. Extract features from the primary curve group and train it to obtain a photovoltaic power anomaly identification and prediction model.
[0010] S3. Obtain current photovoltaic operation data and input the current photovoltaic operation data into the photovoltaic power anomaly identification and prediction model to identify and predict anomalies in the current photovoltaic operation data.
[0011] Preferably, the specific content of constructing a primary curve group from the data in the environment-power mapping group in S2 includes:
[0012] Based on the data categories of environmental data, past environmental data are classified to obtain several environmental parameter category groups, each of which includes the corresponding environmental parameter.
[0013] With time as the X-axis and environmental parameters and power as the Y-axis, establish environmental category curves and power curves;
[0014] By mapping the time points one by one and mapping the environmental curves to the power curves, a set of curves is obtained.
[0015] Preferably, the specific content of extracting features from a set of curves and training a photovoltaic power anomaly identification and prediction model is as follows:
[0016] S201. Perform feature extraction on the primary curve group to obtain the inflection feature, and truncate the primary curve group according to the inflection feature to obtain several secondary curve groups.
[0017] S202. Perform quadratic feature extraction on the quadratic curve group to obtain power anomaly features, and extract data from the quadratic curve group based on the power anomaly features to obtain power anomaly data pairs.
[0018] S203. Remove the power anomaly data groups from the quadratic curve group to obtain missing points, and fill in the missing points to obtain several cubic curve groups.
[0019] S204. Perform three-dimensional feature extraction on the cubic curve set and power anomaly data pair, and train to obtain a photovoltaic power anomaly identification and prediction model.
[0020] Preferably, in step S201, the process of extracting a transition feature from the primary curve group and then truncating the primary curve group based on the transition feature to obtain several secondary curve groups includes:
[0021] The transition features include environmental change features and power transition features corresponding to environmental transition features.
[0022] A feature extraction is performed on a set of curves to obtain the turning point features, and the starting time point of the turning point is marked to obtain the turning point marker point.
[0023] By truncating the curve group according to the turning point, several curve segments are obtained;
[0024] Based on the data categories of environmental transition characteristics, several curve segments are classified to obtain several quadratic curve groups.
[0025] Preferably, the specific content of S204, which involves performing cubic feature extraction on the cubic curve set and training it to obtain the photovoltaic power anomaly identification and prediction model, includes:
[0026] The plurality of cubic curve groups include curve segments of different categories, including curve segments from the same device and curve segments from different devices;
[0027] By splicing the curve segments of the same equipment under different categories according to the original time series, several equipment influence curves are obtained;
[0028] Several environmental impact curves were obtained by splicing curve segments of different equipment under different categories according to the original time series.
[0029] Establish a framework for identifying and predicting photovoltaic power anomalies;
[0030] There are preset time-length cutoff frames. The time-length cutoff frames are used to cut the device influence curve and extract the environmental parameters and power data of the same device within the time-length cutoff frames.
[0031] A photovoltaic power anomaly identification and prediction framework is used to extract features from environmental parameters and power data under the same equipment and train them to obtain the equipment impact index.
[0032] The environmental impact curve is truncated using a time-length truncation box, and environmental parameters and power data for different devices within the time-length truncation box are extracted.
[0033] The photovoltaic power anomaly identification and prediction framework extracts features from environmental parameters and power data under different devices and trains them to obtain an environmental impact index, thereby obtaining a photovoltaic power anomaly identification and prediction model.
[0034] Preferably, the photovoltaic power anomaly identification and prediction framework includes a first convolutional neural module, a second convolutional neural module, a first recurrent neural module, and a second recurrent neural module;
[0035] The first and second convolutional neural modules are constructed in series.
[0036] The first and second circulatory neural modules are constructed in parallel.
[0037] The first convolutional neural module, the second convolutional neural module, the first recurrent neural module, and the second recurrent neural module are constructed in series.
[0038] Preferably, the first convolutional neural module is used to extract features and output a set of quadratic curves.
[0039] The second convolutional neural module is used for secondary feature extraction to output cubic curve sets and power anomaly data pairs;
[0040] The first recurrent neural module is used to extract temporal features from the cubic curve set and train it to perform anomaly recognition and prediction.
[0041] The method is used to extract abnormal features from power anomaly data and train the system to identify specific anomaly categories.
[0042] This application also mentions a distributed photovoltaic power anomaly identification and prediction system based on neural networks, including:
[0043] Data acquisition unit: acquires past photovoltaic operation data, analyzes past operation data to obtain past environmental data and past power data, and maps past environmental data and past power data one-to-one to obtain environmental-power mapping group;
[0044] Model building unit: The data in the environment-power mapping group are used to construct curves to obtain a primary curve group. Features are extracted from the primary curve group and trained to obtain a photovoltaic power anomaly identification and prediction model.
[0045] Identification and prediction unit: acquires current photovoltaic operation data, inputs the current photovoltaic operation data into the photovoltaic power anomaly identification and prediction model to identify and predict anomalies in the current photovoltaic operation data.
[0046] An electronic device is characterized by comprising a memory and a processor, wherein the memory stores a computer program, and the processor, when calling the computer program in the memory, implements the content of a method for identifying and predicting distributed photovoltaic power anomalies based on neural networks.
[0047] A storage medium, characterized in that the storage medium stores computer-executable instructions, which, when loaded and executed by a processor, realize the content of distributed photovoltaic power anomaly identification and prediction based on neural networks.
[0048] In summary, the distributed photovoltaic power anomaly identification and prediction method and system based on neural networks of the present invention, compared with traditional technologies, adopts a series connection of convolutional neural modules and recurrent neural modules, and parallel connection of recurrent neural modules. This improves the prediction accuracy and overcomes the problem of insufficient accuracy due to the inability of short-term or long-term prediction to coordinate their effects. Furthermore, it can identify and judge abnormal situations.
[0049] The technical method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0050] Figure 1 This is a flowchart illustrating the steps of the distributed photovoltaic power anomaly identification and prediction method based on neural networks of the present invention.
[0051] Figure 2 This is a system module diagram of the distributed photovoltaic power anomaly identification and prediction based on neural networks according to the present invention.
[0052] Figure 3 This is a framework diagram for photovoltaic power anomaly identification and prediction in this invention. Detailed Implementation
[0053] The technical method of the present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values of the components and steps described in these embodiments do not limit the scope of this application.
[0054] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.
[0055] Techniques, systems, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, they should be considered part of the instruction manual.
[0056] In all the examples shown and discussed herein, any specific values should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.
[0057] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.
[0058] This invention provides a method for identifying and predicting distributed photovoltaic power anomalies based on neural networks, including the following:
[0059] S1. Obtain past photovoltaic operation data, analyze the past operation data to obtain past environmental data and past power data, and match the past environmental data and past power data one by one to obtain the environment-power mapping group.
[0060] S2. Construct curves from the data in the environment-power mapping group to obtain a primary curve group. Extract features from the primary curve group and train it to obtain a photovoltaic power anomaly identification and prediction model.
[0061] Preferably, the specific content of constructing a primary curve group from the data in the environment-power mapping group in S2 includes:
[0062] Based on the data categories of environmental data, past environmental data are classified to obtain several environmental parameter category groups, each of which includes the corresponding environmental parameter.
[0063] With time as the X-axis and environmental parameters and power as the Y-axis, establish environmental category curves and power curves;
[0064] By mapping the time points one by one and mapping the environmental curves to the power curves, a set of curves is obtained.
[0065] Preferably, the specific content of extracting features from a set of curves and training a photovoltaic power anomaly identification and prediction model is as follows:
[0066] S201. Perform feature extraction on the primary curve group to obtain the inflection feature, and truncate the primary curve group according to the inflection feature to obtain several secondary curve groups.
[0067] S202. Perform quadratic feature extraction on the quadratic curve group to obtain power anomaly features, and extract data from the quadratic curve group based on the power anomaly features to obtain power anomaly data pairs.
[0068] S203. Remove the power anomaly data groups from the quadratic curve group to obtain missing points, and fill in the missing points to obtain several cubic curve groups.
[0069] S204. Perform three-dimensional feature extraction on the cubic curve set and power anomaly data pair, and train to obtain a photovoltaic power anomaly identification and prediction model.
[0070] Preferably, in step S201, the process of extracting a transition feature from the primary curve group and then truncating the primary curve group based on the transition feature to obtain several secondary curve groups includes:
[0071] The transition features include environmental change features and power transition features corresponding to environmental transition features.
[0072] A feature extraction is performed on a set of curves to obtain the turning point features, and the starting time point of the turning point is marked to obtain the turning point marker point.
[0073] Understandably, different weather conditions have varying impacts on photovoltaic (PV) panels. Rainy weather, while blocking sunlight may temporarily reduce power generation, washes away dust from the panel surface, improving long-term efficiency (e.g., power generation may increase by 5%–10% after rain). Similarly, snow cover completely blocks radiation from the panels and requires timely cleaning (as winter snow accumulation can lead to several days of no power generation), but it also helps remove dust. Hail not only affects power output in the immediate future but can also damage the panel glass or backsheet, impacting long-term performance. Therefore, establishing inflection points to classify PV curves can improve the accuracy of later predictions.
[0074] By truncating the curve group according to the turning point, several curve segments are obtained;
[0075] Based on the data categories of environmental transition characteristics, several curve segments are classified to obtain several quadratic curve groups.
[0076] Preferably, the specific content of S204, which involves performing cubic feature extraction on the cubic curve set and training it to obtain the photovoltaic power anomaly identification and prediction model, includes:
[0077] The plurality of cubic curve groups include curve segments of different categories, including curve segments from the same device and curve segments from different devices;
[0078] By splicing the curve segments of the same equipment under different categories according to the original time series, several equipment influence curves are obtained;
[0079] Several environmental impact curves were obtained by splicing curve segments of different equipment under different categories according to the original time series.
[0080] Establish a framework for identifying and predicting photovoltaic power anomalies;
[0081] There are preset time-length cutoff frames. The time-length cutoff frames are used to cut the device influence curve and extract the environmental parameters and power data of the same device within the time-length cutoff frames.
[0082] Understandably, data can be extracted based on different time frame lengths, and different time frame lengths can achieve the fusion of short-term and long-term forecasts.
[0083] A photovoltaic power anomaly identification and prediction framework is used to extract features from environmental parameters and power data under the same equipment and train them to obtain the equipment impact index.
[0084] The environmental impact curve is truncated using a time-length truncation box, and environmental parameters and power data for different devices within the time-length truncation box are extracted.
[0085] The photovoltaic power anomaly identification and prediction framework extracts features from environmental parameters and power data under different devices and trains them to obtain an environmental impact index, thereby obtaining a photovoltaic power anomaly identification and prediction model.
[0086] Preferably, the photovoltaic power anomaly identification and prediction framework includes a first convolutional neural module, a second convolutional neural module, a first recurrent neural module, and a second recurrent neural module;
[0087] The first and second convolutional neural modules are constructed in series.
[0088] The first and second circulatory neural modules are constructed in parallel.
[0089] The first convolutional neural module, the second convolutional neural module, the first recurrent neural module, and the second recurrent neural module are constructed in series.
[0090] Preferably, the first convolutional neural module is used to extract features and output a set of quadratic curves.
[0091] The second convolutional neural module is used for secondary feature extraction to output cubic curve sets and power anomaly data pairs;
[0092] The first recurrent neural module is used to extract temporal features from the cubic curve set and train it to perform anomaly recognition and prediction.
[0093] The method is used to extract abnormal features from power anomaly data and train the system to identify specific anomaly categories.
[0094] It is understandable that some anomalies result in controllable outcomes. Therefore, identifying the types of anomalies helps staff understand the situation and take timely corrective measures.
[0095] S3. Obtain current photovoltaic operation data and input the current photovoltaic operation data into the photovoltaic power anomaly identification and prediction model to identify and predict anomalies in the current photovoltaic operation data.
[0096] Understandably, if the current operation is normal, future operation can be predicted; if abnormalities occur, measures can be taken immediately.
[0097] This application also mentions a distributed photovoltaic power anomaly identification and prediction system based on neural networks, including:
[0098] Data acquisition unit: acquires past photovoltaic operation data, analyzes past operation data to obtain past environmental data and past power data, and maps past environmental data and past power data one-to-one to obtain environmental-power mapping group;
[0099] Model building unit: The data in the environment-power mapping group are used to construct curves to obtain a primary curve group. Features are extracted from the primary curve group and trained to obtain a photovoltaic power anomaly identification and prediction model.
[0100] Identification and prediction unit: acquires current photovoltaic operation data, inputs the current photovoltaic operation data into the photovoltaic power anomaly identification and prediction model to identify and predict anomalies in the current photovoltaic operation data.
[0101] An electronic device is characterized by comprising a memory and a processor, wherein the memory stores a computer program, and the processor, when calling the computer program in the memory, implements the content of a method for identifying and predicting distributed photovoltaic power anomalies based on neural networks.
[0102] A storage medium, characterized in that the storage medium stores computer-executable instructions, which, when loaded and executed by a processor, realize the content of a distributed photovoltaic power anomaly identification and prediction method based on neural networks.
[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical methods of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical methods of the present invention can still be applied in various ways.
[0104] Modifications or equivalent substitutions are made, but these modifications or equivalent substitutions cannot improve the modified technology.
[0105] The method deviates from the spirit and scope of the technical method of this invention.
Claims
1. A method for identifying and predicting distributed photovoltaic power anomalies based on neural networks, characterized in that, Includes the following steps: S1. Obtain past photovoltaic operation data, analyze the past operation data to obtain past environmental data and past power data, and match the past environmental data and past power data one by one to obtain the environment-power mapping group. S2. Construct curves from the data in the environment-power mapping group to obtain a primary curve group. Extract features from the primary curve group and train it to obtain a photovoltaic power anomaly identification and prediction model. S3. Obtain the current photovoltaic operation data, input the current photovoltaic operation data into the photovoltaic power anomaly identification and prediction model to identify and predict anomalies in the current photovoltaic operation data; The specific content of the primary curve group obtained by constructing curves from the data in the environment-power mapping group in S2 includes: Based on the data categories of environmental data, past environmental data are classified to obtain several environmental parameter category groups, each of which includes the corresponding environmental parameter. With time as the X-axis and environmental parameters and power as the Y-axis, establish environmental category curves and power curves; By mapping the time points one by one and mapping the environmental curves to the power curves, a set of curves is obtained. The specific content of extracting features from a set of curves and training a photovoltaic power anomaly identification and prediction model is as follows: S201. Perform feature extraction on the primary curve group to obtain the inflection feature, and truncate the primary curve group according to the inflection feature to obtain several secondary curve groups. S202. Perform quadratic feature extraction on the quadratic curve group to obtain power anomaly features, and extract data from the quadratic curve group based on the power anomaly features to obtain power anomaly data pairs. S203. Remove the power anomaly data groups from the quadratic curve group to obtain missing points, and fill in the missing points to obtain several cubic curve groups. S204. Perform three-dimensional feature extraction on the cubic curve group and power anomaly data pair and train to obtain a photovoltaic power anomaly identification and prediction model. In S201, a feature extraction is performed on a primary curve group to obtain inflection features. Based on these inflection features, the primary curve group is truncated to obtain several secondary curve groups. The specific content includes: The transition features include environmental change features and power transition features corresponding to environmental transition features. A feature extraction is performed on a set of curves to obtain the turning point features, and the starting time point of the turning point is marked to obtain the turning point marker point. By truncating the curve group according to the turning point, several curve segments are obtained; Based on the data categories of environmental transition characteristics, several curve segments are classified to obtain several quadratic curve groups; The specific content of S204, which involves performing cubic feature extraction on a cubic curve set and training a photovoltaic power anomaly identification and prediction model, includes: The plurality of cubic curve groups include curve segments of different categories, including curve segments from the same device and curve segments from different devices; By splicing the curve segments of the same equipment under different categories according to the original time series, several equipment influence curves are obtained; Several environmental impact curves were obtained by splicing curve segments of different equipment under different categories according to the original time series. Establish a framework for identifying and predicting photovoltaic power anomalies; There are preset time-length cutoff frames. The time-length cutoff frames are used to cut the device influence curve and extract the environmental parameters and power data of the same device within the time-length cutoff frames. A photovoltaic power anomaly identification and prediction framework is used to extract features from environmental parameters and power data under the same equipment and train them to obtain the equipment impact index. The environmental impact curve is truncated using a time-length truncation box, and environmental parameters and power data for different devices within the time-length truncation box are extracted. The photovoltaic power anomaly identification and prediction framework extracts features from environmental parameters and power data under different devices and trains them to obtain an environmental impact index, thereby obtaining a photovoltaic power anomaly identification and prediction model.
2. The method for identifying and predicting distributed photovoltaic power anomalies based on neural networks according to claim 1, characterized in that, The photovoltaic power anomaly identification and prediction framework includes a first convolutional neural module, a second convolutional neural module, a first recurrent neural module, and a second recurrent neural module. The first and second convolutional neural modules are constructed in series. The first and second circulatory neural modules are constructed in parallel. The first convolutional neural module, the second convolutional neural module, the first recurrent neural module, and the second recurrent neural module are constructed in series.
3. The method for identifying and predicting distributed photovoltaic power anomalies based on neural networks according to claim 2, characterized in that, The first convolutional neural module is used for a single feature extraction to output a set of quadratic curves; The second convolutional neural module is used for secondary feature extraction to output cubic curve sets and power anomaly data pairs; The first recurrent neural module is used to extract temporal features from the cubic curve set and train it to perform anomaly recognition and prediction. The method is used to extract abnormal features from power anomaly data and train the system to identify specific anomaly categories.
4. A distributed photovoltaic power anomaly identification and prediction system based on neural networks, used to implement the distributed photovoltaic power anomaly identification and prediction method based on neural networks as described in any one of claims 1-3, characterized in that, include: Data acquisition unit: acquires past photovoltaic operation data, analyzes past operation data to obtain past environmental data and past power data, and maps past environmental data and past power data one-to-one to obtain environmental-power mapping group; Model building unit: The data in the environment-power mapping group are used to construct curves to obtain a primary curve group. Features are extracted from the primary curve group and trained to obtain a photovoltaic power anomaly identification and prediction model. Identification and prediction unit: acquires current photovoltaic operation data, inputs the current photovoltaic operation data into the photovoltaic power anomaly identification and prediction model to identify and predict anomalies in the current photovoltaic operation data.
5. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor, when calling the computer program in the memory, implements the content of the distributed photovoltaic power anomaly identification and prediction method based on neural networks as described in any one of claims 1 to 3.
6. A storage medium, characterized in that, The storage medium stores computer-executable instructions, which, when loaded and executed by a processor, implement the method for identifying and predicting distributed photovoltaic power anomalies based on neural networks as described in any one of claims 1 to 3.