The rapid expansion of photovoltaic power generation has fundamentally reshaped the global energy landscape, driven by the dual imperatives of reducing carbon emissions and achieving energy sustainability. Solar energy, as one of the most abundant and accessible renewable resources, has attracted substantial investment and technological innovation across the world. Changqing Oilfield, in alignment with the broader global energy transition, has strategically introduced photovoltaic power projects to reduce dependence on fossil fuels and support low-carbon operational objectives. Within this framework, the solar inverter serves as the pivotal component of the photovoltaic power generation system, acting as the interface between the direct current output of solar panels and the alternating current required by the electrical grid. The output current of the solar inverter directly reflects its operational status, and any abnormal fluctuation can significantly affect power generation efficiency, equipment safety, and grid stability. Therefore, precise monitoring of the current parameters of solar inverters and the establishment of an effective anomaly detection mechanism are essential for ensuring the reliable operation of photovoltaic systems.
The operational environment of solar inverters is inherently complex and variable. The current output is influenced by numerous factors, including weather conditions, photovoltaic panel performance, equipment degradation, and grid disturbances. Traditional anomaly detection methods, often relying on fixed thresholds or simple statistical models, frequently suffer from high false positive rates and missed detections, especially when confronted with non-stationary and highly volatile time series data. To address these limitations, this thesis leverages data mining techniques to investigate and develop robust current anomaly detection methods for solar inverters. The core research is structured around three principal areas: data preprocessing and clustering to construct weather-specific datasets; the development and evaluation of two deep learning-based anomaly detection models, namely VMD-BiGRU and FFT-AE; and the design and implementation of a comprehensive detection system for photovoltaic inverter currents.
1 Data Preprocessing and Clustering of Solar Inverters
The foundation of any data-driven approach lies in the quality and structure of the underlying data. For solar inverters, the raw data collected from sensors, gateways, and cloud platforms often contains noise, missing values, outliers, and redundant attributes that can compromise model performance. To ensure the integrity and reliability of subsequent analysis and anomaly detection, a rigorous data preprocessing pipeline was designed and implemented.
1.1 Data Acquisition and Cleaning
The data acquisition architecture for the photovoltaic power station is structured into three layers: the device layer (solar inverters), the communication layer (4G gateways), and the application layer (cloud platform). The solar inverter data is collected via an RS485 connection using the Modbus RTU protocol. The gateway processes the data and transmits it over a 4G wireless network to the cloud platform, which decodes and stores the data for analysis.
Data cleaning is a critical first step in preparing the dataset for analysis. The following operations were performed to enhance data quality:
- Invalid Attribute Removal: A significant portion of the collected data attributes, such as voltage imbalance, over-temperature status, and short-circuit indicators, contained numerous zero values or irrelevant status flags for current anomaly detection. These attributes were removed to reduce dimensionality and noise, leaving 13 relevant data items, including A/B/C phase currents and voltages, efficiency, frequency, temperature, power, and power factor.
- Data Resampling: The raw data was recorded at an interval of every 3 seconds. To reduce data volume and redundancy while preserving essential temporal patterns, the data was down-sampled to 10-minute intervals. Furthermore, since the solar inverter primarily generates power between 05:30 and 20:30, records outside this time window, which typically contain null current values, were filtered out to focus on periods of active operation.
- Missing Value Imputation: Data completeness is essential for model training. Linear interpolation was selected as the primary method for filling missing values. This method assumes a linear relationship between data points, and when combined with a moving average smoothing technique, it effectively handles the fluctuations in the current data. The moving average calculation is defined as follows:
$$SMA_t = \frac{x_1 + x_2 + \cdots + x_n}{n}$$
where $x_i$ is the data point and $n$ is the time period. - Outlier Removal: Anomalous values arising from sensor malfunctions or transmission errors were identified and treated as missing values, subsequently filled using the interpolation method to maintain a consistent dataset.
1.2 Feature Selection Based on Principal Component Analysis
Feature selection aims to reduce model complexity and identify critical features for anomaly detection. Principal Component Analysis (PCA) is a widely used technique for dimensionality reduction that transforms a set of correlated variables into a smaller number of uncorrelated components. Given the high-dimensional and correlated nature of the solar inverter data, PCA was selected as the appropriate method for feature selection.
The PCA process begins with standardizing the data matrix. For an $n \times p$ matrix $X$, the standardized values are computed using:
$$\bar{x}_j = \frac{1}{n} \sum_{i=1}^{n} x_{ij}, \quad S_j = \sqrt{\frac{\sum_{i=1}^{n} (x_{ij} – \bar{x}_j)^2}{n-1}}, \quad X_{ij} = \frac{x_{ij} – \bar{x}_j}{S_j}$$
After standardizing the data, the correlation coefficient matrix is calculated to determine the relationships between attributes and to derive the eigenvalues and eigenvectors. The variance explained by each principal component $b_i$ and the cumulative variance $\alpha_i$ are determined by:
$$b_i = \frac{\lambda_i}{\sum_{k=1}^{p} \lambda_k}, \quad \alpha_i = \frac{\sum_{k=1}^{i} \lambda_k}{\sum_{k=1}^{p} \lambda_k}$$
In this study, PCA was applied to the inverter data collected on March 30, 2023, excluding the current and date attributes. The analysis of the variance explained ratio indicated that selecting five principal components was sufficient, as they achieved a cumulative contribution rate of over 84%. The principal component analysis results are summarized in the following table:
| Attribute | Contribution Rate | Cumulative Contribution Rate |
|---|---|---|
| nbq_f (Grid Frequency) | 0.249938 | 0.249938 |
| nbq_p (Active Power) | 0.243829 | 0.493766 |
| nbq_t (Inverter Temperature) | 0.157906 | 0.651672 |
| nbq_pf (Power Factor) | 0.114502 | 0.766174 |
| nbq_ub (B-Phase Voltage) | 0.073489 | 0.839663 |
Through this process, five key feature attributes were selected to complement the target attribute of inverter current. The final set of features used for the dataset includes grid frequency (Hz), active power (kW), inverter temperature (°C), power factor, B-phase voltage (V), and the A-phase current (A) of the solar inverter.
1.3 Analysis of Characteristic Indicators
Understanding the operational characteristics of the selected features is crucial for building effective anomaly detection models. The analysis revealed distinctive patterns for each indicator of the solar inverter:
| Indicator | Normal Operating Range | Characteristics |
|---|---|---|
| Grid Frequency (nbq_f) | 49.5 – 50.5 Hz | Stable around 50 Hz, fluctuations indicate grid instability. |
| Active Power (nbq_p) | 80% – 100% of rated power | Highly correlated with solar irradiance and current. |
| Inverter Temperature (nbq_t) | 5 – 45 °C | Varies with ambient temperature and load; excessive values indicate faults. |
| Power Factor (nbq_pf) | 0.8 – 1 | Values near 1 indicate high efficiency. |
| B-Phase Voltage (nbq_ub) | 220 – 250 V | Relatively stable under normal operating conditions. |
| A-Phase Current (nbq_ia) | 0 – 40 A | Fluctuates with solar irradiance, showing a daily bell-shaped curve. |
The analysis of the inverter current in solar inverters revealed that its daily profile generally shows a rising trend from morning to noon, reaching a peak around midday, followed by a declining trend toward the evening. This pattern reflects the relationship between photovoltaic power generation and meteorological factors. However, the intrinsic volatility of weather leads to distinct current patterns under different conditions. Observations allowed for the categorization of inverter current into three primary weather types: sunny, cloudy, and rainy/overcast. Sunny days show a smooth, gaussian-like current curve; cloudy days exhibit high volatility and significant amplitude differences; and rainy/overcast days are characterized by stable yet suppressed current values.
1.4 Construction of Weather-Specific Datasets via DTW-K-means
To improve the accuracy and robustness of anomaly detection for solar inverters, it is advantageous to build separate models for distinct weather scenarios. K-means clustering is a simple and efficient algorithm for this purpose. However, the standard K-means algorithm, which relies on Euclidean distance, is often inadequate for time series data due to its sensitivity to distortions in the time axis. To address this limitation, the Dynamic Time Warping (DTW) distance metric was adopted to replace the Euclidean distance in the K-means algorithm, creating a DTW-K-means clustering method.
DTW is an algorithm that measures the similarity between two temporal sequences by finding an optimal alignment between their data points. The core objective of DTW is to minimize the cumulative distance along the optimal warping path. For two sequences $Q$ and $S$, the DTW distance is defined as:
$$DTW(Q, S) = D(n,m) = w_{nm} + \min\{D(n-1,m), D(n,m-1), D(n-1,m-1)\}$$
In this study, the DTW-K-means algorithm was used to cluster the inverter current data into three categories, representing sunny, cloudy, and rainy/overcast weather conditions. The clustering was based on statistical features of the current, including mean, maximum, minimum, standard deviation, and variance. A comparison of the clustering results from the standard K-means and the DTW-K-means algorithms demonstrated that the DTW-based approach yields more distinct and well-separated clusters, effectively distinguishing between the different weather patterns. This optimized clustering method provides a solid foundation for training specialized anomaly detection models, leading to improved detection performance across various meteorological conditions.
2 Deep Learning Models for Inverter Current Anomaly Detection
In this thesis, two sophisticated deep learning-based anomaly detection models were developed to identify abnormal currents in solar inverters: a prediction-based model named VMD-BiGRU and a reconstruction-based model named FFT-AE. Each model employs a unique architectural strategy to capture the complex patterns in the data and identify deviations from normal behavior.
2.1 The VMD-BiGRU Anomaly Detection Model
The VMD-BiGRU prediction model combines Variational Mode Decomposition (VMD) with a Bidirectional Gated Recurrent Unit (BiGRU) network. This hybrid architecture is designed to enhance the quality of the input data and effectively capture temporal dependencies in both forward and backward directions.
Variational Mode Decomposition (VMD): VMD is an adaptive, non-recursive signal processing technique that decomposes a complex input signal into a series of band-limited intrinsic mode functions (IMFs). The optimal decomposition aims to minimize the sum of the bandwidths of the modes while ensuring that the sum of all modes reconstructs the original signal. The constrained variational problem is expressed as:
$$\min_{\{u_k\},\{\omega_k\}} \left\{ \sum_{k=1}^{K} \left\| \partial_t \left[ \left( \delta(t) + \frac{j}{\pi t} \right) \times u_k(t) \right] e^{-j\omega_k t} \right\|_2^2 \right\} \quad \text{s.t.} \quad \sum_{k=1}^{K} u_k(t) = f(t)$$
In this study, VMD was applied to the inverter current data, decomposing it into a set of IMFs. The decomposed components, such as IMF1, IMF2, and IMF3, were then added to the dataset as new features, enriching the information available for model training. The decomposition effectively captures the underlying trends and noise components of the current signal, thereby improving the model’s ability to detect subtle anomalies in solar inverters.
Bidirectional Gated Recurrent Unit (BiGRU): GRU is a variant of Recurrent Neural Networks (RNN) designed to address the vanishing gradient problem. It employs two gate mechanisms: an update gate and a reset gate. The update gate $z_t$ controls how much of the past information needs to be passed along to the future, while the reset gate $r_t$ determines how much of the past information to forget. The GRU computations at time step $t$ are described by:
$$z_t = \sigma(W_z \cdot [h_{t-1}, x_t])$$
$$r_t = \sigma(W_r \cdot [h_{t-1}, x_t])$$
$$\hat{h}_t = \tanh(W \cdot [r_t \odot h_{t-1}, x_t])$$
$$h_t = (1 – z_t) \odot h_{t-1} + z_t \odot \hat{h}_t$$
To capture both past and future context, BiGRU processes the sequence in two directions: a forward GRU that reads the sequence from start to end and a backward GRU that reads from end to start. The final hidden state $h_t$ at each time step is a combination of the forward hidden state $\overrightarrow{h}_t$ and the backward hidden state $\overleftarrow{h}_t$:
$$h_t = \omega_t \overrightarrow{h}_t + v_t \overleftarrow{h}_t + b_t$$
The VMD-BiGRU anomaly detection workflow for solar inverters comprises the following steps:
- Data Preprocessing: Apply data cleaning and feature selection to the raw inverter data.
- Data Clustering: Use DTW-K-means to create three weather-specific datasets: sunny, cloudy, and rainy.
- Feature Decomposition: Apply VMD to the inverter current component in each weather dataset to generate new feature channels.
- Model Training and Prediction: Train the BiGRU network on the augmented dataset and generate current predictions.
- Residual Calculation: Compute the residuals between the predicted and actual current values.
- Anomaly Detection: Apply the z-score method to the residuals to detect anomalies. A threshold of 2.5 standard deviations was used in the experiments to flag data points as abnormal.
The z-score for a residual value $x$ is calculated as:
$$z\text{-}score = \frac{x – \mu}{\sigma}$$
where $\mu$ is the mean of the residual data and $\sigma$ is its standard deviation. A high absolute z-score indicates a significant deviation, which is likely an anomaly in solar inverters.
2.2 The FFT-AE Anomaly Detection Model
The FFT-AE reconstruction model integrates the Fast Fourier Transform (FFT) with a traditional Autoencoder (AE) network. This model operates by transforming the time-domain current signal into the frequency domain, learning the essential frequency features, and then reconstructing the original time-domain signal. A high reconstruction error indicates an anomaly.
Fast Fourier Transform (FFT): The Discrete Fourier Transform (DFT) is a fundamental tool for frequency analysis of discretized signals. Given a sequence of $N$ samples $x[n]$, the DFT is defined as:
$$X[k] = \sum_{n=0}^{N-1} x[n] e^{-j \frac{2\pi}{N} kn}$$
The FFT is an efficient algorithm to compute the DFT, reducing the computational complexity from $O(N^2)$ to $O(N log N)$. By decomposing the DFT into a combination of smaller transforms, the FFT enables rapid analysis of the frequency components. For solar inverters, this analysis helps in identifying the significant frequency components of the current signal, effectively filtering out noise and irrelevant information. The FFT divides the input sequence into even and odd indexed parts to achieve this efficiency.
Autoencoder (AE): An autoencoder is a type of neural network trained to learn a compressed representation of the input data. It consists of an encoder, which maps the high-dimensional input to a lower-dimensional latent space, and a decoder, which attempts to reconstruct the original input from this compressed representation. The structure can be represented as:
$$z = f(w^{(1)}x + b^{(1)})$$
$$x’ = f(w^{(2)}z + b^{(2)})$$
The training process involves minimizing the difference between the original input and the reconstruction, typically using the Mean Squared Error (MSE) as the loss function:
$$MSE = \frac{1}{n} \sum_{i=1}^{n} (x_i – x’_i)^2$$
Once trained on normal data, the autoencoder becomes proficient at reconstructing “normal” patterns. If an input sample contains an anomaly, its reconstruction error will be high, as a solar inverter operating under abnormal conditions will have frequency characteristics that deviate from the learned normal patterns.
The FFT-AE anomaly detection workflow for solar inverters is as follows:
- Data Preprocessing and Clustering: Clean the collected current data and cluster it into the three weather-specific datasets using the DTW-K-means algorithm.
- Time-to-Frequency Conversion: Apply the FFT to the time-domain current data to convert it into the frequency domain. Retain the top 200 most significant frequency components to represent the signal’s key characteristics.
- Autoencoder Processing: Feed the frequency-domain features into the autoencoder. The encoder compresses the input to a low-dimensional representation, and the decoder attempts to reconstruct the input frequency features from this representation.
- Frequency-to-Time Conversion: Apply the Inverse Fast Fourier Transform (IFFT) to the reconstructed frequency features to obtain the reconstructed time-domain current signal.
- Anomaly Identification: Compute the reconstruction error between the original time-domain signal and the reconstructed signal. Use the z-score method to set a threshold (3 standard deviations) to classify data points with high reconstruction errors as anomalies in solar inverters.
3 Experimental Evaluation and Performance Analysis
3.1 Experimental Setup and Data Description
All experiments were conducted in a consistent software and hardware environment to ensure fairness and comparability of results. The specific experimental environment is detailed below:
| Name | Configuration/Version |
|---|---|
| Operating System | Windows 11 |
| CPU | 12th Gen Intel(R) Core(TM) i5-12500H |
| GPU | NVIDIA GeForce RTX 3050 Laptop GPU |
| Development Environment | PyCharm 2022.2.3 |
| Deep Learning Framework | TensorFlow 2.15.0 |
| Programming Language | Python 3.9.13 |
| Machine Learning Library | Scikit-learn 1.0.2 |
| Numeric Computing | Numpy 1.21.5, Pandas 1.4.4 |
The experimental data was sourced from the actual operational data of a photovoltaic power station in Changqing Oilfield, spanning from March 22, 2023, to September 30, 2023. After resampling, the dataset comprised 17,509 records. Following the DTW-K-means clustering, the data was partitioned into three weather-specific subsets. The sunny dataset contained 11,640 records (8,148 for training and 3,492 for testing). The cloudy dataset contained 4,005 records (2,800 for training and 1,205 for testing). The rainy/overcast dataset contained 1,864 records (1,300 for training and 564 for testing).
3.2 Model Parameter Configuration
The parameters for both models were carefully tuned through multiple iterations of experiments to achieve optimal performance. The specific parameter settings for the VMD-BiGRU and FFT-AE models for solar inverters are presented below:
| Parameter (VMD-BiGRU) | Value | Parameter (FFT-AE) | Value |
|---|---|---|---|
| VMD Decomposition Modes | 3 | Learning Rate | 0.001 |
| Time Step | 90 | Epoch | 100 |
| Learning Rate | 0.001 | Optimizer | Adam |
| Batch Size | 32 | Loss Function | MSE |
| Epoch | 100 | Residual Threshold Factor | 3 |
| Optimizer | Adam | FFT Retained Frequencies | 200 |
| Loss Function | MSE | AE Input Dimension | 200 |
| Residual Threshold Factor | 2.5 |
3.3 Comparative Results and Discussion
To validate the effectiveness of the proposed methodology, a series of experiments were conducted. First, the clustering process was validated by performing anomaly detection on both the unclustered and clustered datasets using traditional statistical methods: fixed threshold, boxplot, and z-score. The results are shown in the following table:
| Category | Indicator | Actual Anomalies | Fixed Threshold | Boxplot | Z-score |
|---|---|---|---|---|---|
| Before Clustering | Entire Current Dataset | 347 | 38 | 48 | 73 |
| After Clustering | Sunny Current | 230 | 24 | 13 | 21 |
| Cloudy Current | 76 | 10 | 74 | 90 | |
| Rainy/Overcast Current | 41 | 22 | 95 | 34 | |
| Total | 347 | 56 | 182 | 95 |
The comparison between clustered and unclustered results clearly shows that the clustering method significantly enhances the sensitivity of traditional anomaly detection methods. However, these methods still exhibit high false positive and false negative rates, highlighting the need for more sophisticated deep learning models for solar inverters.
Subsequently, the performance of the VMD-BiGRU and FFT-AE models was evaluated on the clustered datasets. The detected anomalies for each model are presented below:
| Detection Indicator | Actual Anomalies | VMD-BiGRU | FFT-AE |
|---|---|---|---|
| Sunny Current | 53 | 57 | 51 |
| Cloudy Current | 29 | 33 | 30 |
| Rainy/Overcast Current | 32 | 27 | 31 |
| Total | 114 | 117 | 112 |
3.4 Model Performance Evaluation Metrics
To quantitatively evaluate and compare the performance of the two deep learning models, the Confusion Matrix was used, which categorizes detection outcomes into four types: True Positive (TP), False Negative (FN), False Positive (FP), and True Negative (TN). Based on these values, several key performance metrics were calculated:
- Precision: The proportion of correctly predicted anomalies out of all predicted anomalies.
$$Precision = \frac{TP}{TP + FP}$$ - Recall: The proportion of actual anomalies that were correctly identified.
$$Recall = \frac{TP}{TP + FN}$$ - Accuracy: The proportion of all correct predictions (both anomalies and normals) out of the total samples.
$$Accuracy = \frac{TP + TN}{TP + TN + FP + FN}$$ - F1 Score: The harmonic mean of precision and recall, providing a balanced measure of the model’s performance.
$$F1Score = \frac{2 \times Precision \times Recall}{Precision + Recall}$$
The evaluation metrics for both models were calculated on each weather-specific dataset, and the results are presented in the following tables:
| Metric (Sunny) | VMD-BiGRU (%) | FFT-AE (%) |
|---|---|---|
| F1 Score | 93.09 | 94.30 |
| Accuracy | 93.50 | 96.10 |
| Recall | 92.81 | 96.11 |
| Metric (Cloudy) | VMD-BiGRU (%) | FFT-AE (%) |
|---|---|---|
| F1 Score | 90.53 | 95.10 |
| Accuracy | 91.68 | 96.57 |
| Recall | 91.10 | 95.73 |
| Metric (Rainy) | VMD-BiGRU (%) | FFT-AE (%) |
|---|---|---|
| F1 Score | 93.51 | 97.50 |
| Accuracy | 95.73 | 98.60 |
| Recall | 94.42 | 97.25 |
The comprehensive evaluation metrics, averaged across all weather conditions, are shown below:
| Metric (Overall Average) | VMD-BiGRU (%) | FFT-AE (%) |
|---|---|---|
| F1 Score | 92.37 | 95.63 |
| Accuracy | 93.64 | 97.09 |
| Recall | 92.78 | 96.36 |
The experimental results demonstrate that both deep learning models perform exceptionally well in detecting anomalies in the current of solar inverters, achieving F1 scores consistently above 90%. The FFT-AE model consistently outperforms the VMD-BiGRU model across all metrics and weather conditions. This performance difference can be attributed to the algorithmic mechanisms of the two models. The VMD-BiGRU model relies on historical time series data for prediction, which can introduce a lag in learning the continuously changing characteristics of solar inverter currents, leading to accumulated prediction errors. In contrast, the FFT-AE model reconstructs data based on the frequency characteristics of the current time window, mitigating the issue of temporal dependency. Furthermore, both models achieve their best performance on the rainy/overcast dataset due to the relatively stable current variations, which are easier for both models to handle. Overall, the FFT-AE model proves to be a more reliable and accurate solution for current anomaly detection in solar inverters.
4 Design and Implementation of the Photovoltaic Inverter Current Anomaly Detection System
To bridge the gap between theoretical research and practical application, a comprehensive photovoltaic inverter current anomaly detection system was designed and implemented in this thesis. The system architecture adopts a straightforward three-layer design pattern, encompassing the data storage layer, business logic layer, and presentation layer. The data storage layer is responsible for maintaining raw data and system logs, the business logic layer handles the core detection algorithms and application logic, and the presentation layer offers an intuitive user interface. The system is built upon the Springboot and Vue frameworks, utilizing MySQL for persistent data storage, Redis for caching, and Echarts for data visualization.
4.1 Requirements and Functional Module Analysis
A thorough requirements analysis was conducted based on the actual working conditions of small-scale photovoltaic power stations. The primary functional requirements of the system were identified and structured into four core modules:
- System Management: This module includes user management, role management, department management, menu management, and log management. It ensures secure access control and provides an audit trail for user and system operations. By defining different roles and allocating specific permissions, the system guarantees that users can only access the functions necessary for their duties.
- Data Management: This module is centered around the management of inverter information and its collected operational data. It supports queries for historical data on current, voltage, power, and temperature, as well as the visualization of voltage and current trends, enabling users to monitor the real-time status of solar inverters effectively.
- Prediction and Anomaly Detection: This is the core functional module of the system, integrating the trained VMD-BiGRU and FFT-AE models. It allows users to generate current forecasts and execute anomaly detection tasks, with the results being clearly displayed and accessible for analysis.
- Operation and Maintenance Record Management: This module is divided into two parts: alarm records and repair records. It is used to log any alarms triggered by anomalies and to document the maintenance history of the solar inverters, providing valuable data for preventive maintenance and fault analysis.
4.2 Database Design
Following the requirements analysis, a detailed database design was developed, beginning with the conceptual model. This involved identifying the principal entities within the system and their relationships. Entities were categorized into two main groups: system management entities and inverter data management entities. Based on the Entity-Relationship (E-R) diagrams, the logical structure for the database was designed. Key database tables include:
| Table Name | Description | Key Fields |
|---|---|---|
| sys_user | Stores user account information. | user_id, dept_id, user_name, password, email, phone_number |
| sys_role | Stores role definitions and permissions. | role_id, role_name, role_key, status |
| sys_menu | Manages the system menu structure and routing. | menu_id, menu_name, parent_id, component, perms |
| sys_dept | Manages the organizational structure of departments. | dept_id, parent_id, dept_name, leader, phone |
| inverter_info | Stores basic technical information about the solar inverters. | inverter_id, model, manufacture, install_date |
| inverter_cjdata | Stores the collected operational data from solar inverters. | cj_id, date, inverter_id, nbq_ua, nbq_ub, nbq_uc, nbq_ia, nbq_ib, nbq_ic, nbq_p, nbq_t |
| alarm_info | Stores alarm records triggered by anomalies. | alarm_id, alarm_time, inverter_id, alarm_type, level |
| repair_info | Stores the maintenance and repair history for solar inverters. | repair_id, inverter_id, repair_user, repair_date, repair_content |
| predict_result | Stores the results of the current prediction model. | inverter_id, date, act_current, pre_current |
| anomaly_result | Stores the results of the anomaly detection models. | inverter_id, date, act_current, vmd-bigru, fft-ae |
4.3 System Implementation and Testing
The system was fully implemented according to the designed architecture and database schema. The user interface was developed using Vue and Element UI, offering a clean, responsive, and user-friendly experience. The backend services, based on Springboot, handle business logic, data processing, and interactions with the deep learning models. Key functional interfaces for managing solar inverters and performing anomaly detection were developed and rigorously tested. The system successfully provides essential features, including the real-time monitoring of inverter currents, comprehensive data management, and the ability to identify anomalies using the two pre-trained deep learning models.
System testing was a critical phase to ensure the stability and reliability of the application. A series of functional test cases were designed and executed for each core module, covering operations like user login, CRUD operations for user and role management, data querying and monitoring, the execution of prediction and anomaly detection tasks, and the management of alarms and repair records. All test cases were executed successfully, and the observed results matched the expected behavioral outcomes, confirming that the system meets its specified functional requirements and is ready for real-world deployment in the operational management of solar inverters.
5 Conclusions and Future Work
This thesis has presented a comprehensive study on data mining-based current anomaly detection methods for solar inverters, addressing key challenges in data quality, accurate detection, and practical system deployment. The primary contributions and findings of this research can be summarized as follows:
First, a rigorous data preprocessing pipeline was established to guarantee the quality and consistency of the dataset. Invalid attributes were eliminated, data was resampled to an appropriate frequency, and missing values were filled using a combination of linear interpolation and moving average smoothing. Principal Component Analysis was effectively employed for feature selection, resulting in a set of six key indicators that greatly improve the efficiency and accuracy of anomaly detection for solar inverters.
Second, the DTW-K-means clustering algorithm was proposed to construct weather-specific datasets. By substituting the traditional Euclidean distance with the DTW distance, the clustering algorithm successfully captured the temporal and nonlinear fluctuations in inverter current data. The resulting clusters, corresponding to sunny, cloudy, and rainy/overcast conditions, provide a solid basis for training specialized and more accurate anomaly detection models for solar inverters.
Third, two sophisticated deep learning models, VMD-BiGRU and FFT-AE, were successfully developed and validated for inverter current anomaly detection. The experimental results demonstrated that both models achieve high detection accuracy and recall, exceeding 90% under various conditions. The FFT-AE model, in particular, exhibited superior performance with an overall accuracy of 97.09% and an F1-score of 95.63%, highlighting its effectiveness in reconstructing normal current patterns and identifying deviations in solar inverters.
Finally, a comprehensive photovoltaic inverter current anomaly detection system was designed and implemented based on the Springboot and Vue frameworks. The system offers practical functionality including data management, system administration, current prediction, anomaly detection, and maintenance record tracking, providing a valuable tool for photovoltaic power station operations.
For future research, several directions can be considered to further enhance the capabilities of inverter current anomaly detection. The prediction accuracy of the VMD-BiGRU model could be improved by incorporating attention mechanisms or other advanced sequence modeling techniques to better capture complex data patterns. Additionally, the current study focuses on single-phase current analysis; future work could explore multi-dimensional correlation analysis across A, B, and C-phase currents to provide a more comprehensive and robust assessment of the operational state of solar inverters. Moreover, while the models are effective at detecting anomalies, research into the root causes of these anomalies, using interpretability techniques or additional context data, would be highly beneficial for proactive maintenance and prevention strategies in the operation of solar inverters.