FAULT DETECTION IN SOLAR POWER PLANTS USING PREDICTIVE ANALYTIC TECHNIQUES

FAULT DETECTION IN SOLAR POWER PLANTS USING PREDICTIVE ANALYTIC TECHNIQUES DR. V. ANANTHA NATARAJAN 1, DR. K. SUNEETHA 2, DR. GIRISHA L 3, DR. MOTI LAL 1RINAWA 4 1 Associate Professor, Dept of CSE, Sree Vidyanikethan Engineering College Tirupati, AP, India. v.ananth.satyam@gmail.com 2 Professor, School of CS & IT, JAIN(Deemed-to-be University), Bangalore, India. umasuni.k@gmail.com 3 Associate Professor, Dept of Mechanical Engineering, PES Institute of Technology and Management, Sagara Road, SHIVAMOGGA, KA, India. girinaik@gmai.com 4 Assistant Professor, Dept of Mechanical, Engineering, Government Engineering College, Jhalawar, Rajasthan, India. Riinawa.iitd@gmail.com ABSTRACT In general solar power is generated by converting the energy available in the sunlight in to electricity using photovoltaics semi conducting materials, or concentrated solar power. This research study focuses on photovoltaic based solar power generation where the photovoltaic effects are used for generating DC which is then converted in to AC power by using the inverters. The aim of this research work is to develop methods to predict the near future power generation (energy forecast) and perform fault identification of an equipment. The prediction problem is solved using a linear and non-linear regression algorithm. The performance of the support vector regression, a non-linear algorithm was observed to be superior in certain cases. The regression algorithms are trained using the weather data captured using sensors as the independent variable and the generated power as the dependent variable. Keyboards: fault identification, energy forecast, support vector regression, weather data, photovoltaic I. INTRODUCTION Solar power has become an important source of renewable energy and alternate energy source. The radiations emitted by the sun can be transformed directly in to electric power by photovoltaic cells. Due to photovoltaic effect a small amount of electric power is produced as the light source from the sun directly falls on the junction between two semi-conductor materials. The power generated from the single cell is small and for generating electric power in range of multiple megawatts a series of photovoltaic cells are arranged in a grid fashion (array of solar panels). The efficiency of a photovoltaic cell is only 15-20% and different solutions have been proposed to increase its efficiency. When sunlight falls on a photovoltaic cell due to photoelectric effect an electron is freed and it makes the electrons to travel to the external circuit thus power is supplied to the load. The properties and characteristics of the semiconductors decide the electric flow which in turn depends on the intensity of the light falling on the surface of the photovoltaic cells. Especially in a tropical country like India solar power has been largely generated, as of Nov, 2020 the total capacity of solar power generation in India is around 36.9 GigaWatts. The efficiency of the solar power production can be increased when certain issues in the photovoltaic based power generation is addressed. Often in solar power generation due to electrical and mechanical faults in the photovoltaic cells the overall power production is affected. There is need for methods to maintain the performance and lower the downtime with anearly fault detection. To achieve the above stated objective of having low downtime and increased efficiency the malfunctions in the cells have to be detected and communicated to the grid operator for necessary action. Statistical techniques and machine learning algorithms can be used analyze and interpret the faults from a series of data observed in the solar plant. Reliability of a power generation plant is a critical factor which needs to be given more importance and efficiently handle issues including ground faults, short circuits, soiling, and shading. In general few of the issues www.turkjphysiotherrehabil.org 2500

or faults remain undetected for longer duration and hinder the power production. This at time may lead threat to life and dangerous to the surrounding equipment. The ground faults occurring in the system can be detected using inverters but whereas soiling or short circuits may remain undetected. With less effort and time the IV data can be measured at each solar panel where I and V have high correlation. The IV curve can be modeled using a single diode model by considering the temperature, solar irradiance, short-circuit current denoted as I sc and open circuit voltage V oc. Each solar panel has a upper limit in the power generation which is termed as Maximum Power Point (MPP) and the solar panel faults can be detected by approximating the MPPS and estimating the variation between the approximated and observed values. This paper builds regression model for accurately predicting the future power generation based on the measured temperature and solar irradiation values. Initially using few exploratory analysis techniques the characteristics and nature of the data variables were explored to identify their dependency between the other variables. The proposed approach is a time series forecasting of power generation using the local weather estimates and using anomaly detection techniques the fault in the equipment can be identified. II. RELATED WORK In few researches to detect the solar panel faults, statistical methods have been employed [1]. Some other research attempts have explored the advantage of using neural architecture for detecting the faults [2-5]. Till now there is gap between the proposed approaches for fault detection and the actual requirement. A generic method or approach is essential for detecting and recognizing the solar panel faults when there is little deviation from the usual power generation. Clustering algorithms were also used for detecting the faults in solar panel [6-8]. This paper shall explore the advantage of using machine learning techniques in detecting the solar panel faults using the observed data and weather conditions. In the existing power production system the average mean time to repair is 19 days and the power losses can be reduced when the average mean time to repair is reduced. A fault detection system must be able to accurately discriminate between the normal and anomalous condition prevailing in the solar panel. When an anomaly is detected it should recognize the cause and category of the faults. The major concern is the unavailability of large volume of labeled solar panel data describing the various categories of faults. The circulation of the generated electric current may be interrupted at time due to disconnection in the network causing open-circuit faults. This leads to a major impact on the power production and the severity of the impact depends on the location of the fault and network connectivity pattern in the system. Because of a low impendence path appearing in the system the short circuit faults may arise. In conventional approaches for detecting the fault in solar system, the power generation system is modeled as a statistical or machine learning model and their results are compared with the data extracted at real time to detect any anomalies when the difference between the model output and the real time data is larger than some predefined threshold value [9][19]. In case of statistical methods based detection using a multivariate exponential weighted moving average the limitations of the existing approaches were resolved. Residual values of DC current, power generated, voltage, temperature and the solar irradiance were used as input to the statistical model. Using the residual the difference between the measured values and the predictions can be made and used for detecting the fault in the system. The generalization ability of the proposed approach in detecting the other faults was not explored by the authors. After detecting a fault the category of faults can be identified using an artificial intelligence model particularly a machine learning classifier [10][17]. Using artificial neural network the operation of solar power generation system is categorized as normal, degradation, short-circuit, and shadowing in [11][15]. The neural network is trained and tested with the data generated from multiple simulations of the system. In another work a stage approach was proposed for detecting and recognizing the category of fault in the solar power generation system. For detecting the fault occurring in the system the power generated is compared with output from the mathematical model of the system. If the difference is larger than a threshold or reference value then a fault is confirmed and the fault is classified using a multilayer feed-forward neural network. The major drawback is that the system is trained with the simulated data and tested with the real data captured from the voltage vs current curve. Using the same kind of approach faults were detected and classified in [12][16][18] using the simulation data. Recently two other research articles [13, 14] presented a two stage detection architecture using non-linear auto regressive models for estimating the power generation at different atmospheric condition. At the second stage, fuzzy inference models were used to compare the estimated value to the actual generated power for classifying the fault in to one of the fault categories (shadowing, short and open circuit) considered. www.turkjphysiotherrehabil.org 2501

DATASET DESCRIPTION The dataset used for experimental analysis is collected from Kaggle, a public data platform. The dataset was recorded at 02 solar power plants located in India and the collected data include power generation data and weather parameters data. Each inverter is connected to a series of solar panel and the solar power generation data is collected at the inverter level and the weather data is collected from the sensor located at the respective power plant. From the collected it is possible to identify an anomaly in the system and faulty or underperforming equipment. The power generation data includes the following information 1. DATE_TIME: Date and time for each observation. Observations recorded at 15 minute intervals. 2. PLANT_ID: this will be common for the entire file 3. SOURCE_KEY: Source key in this file stands for the inverter id. 4. DC_POWER: Amount of DC power generated by the inverter (source_key) in this 15 minute interval. Units - kw. 5. AC_POWER: Amount of AC power generated by the inverter (source_key) in this 15 minute interval. Units - kw. 6. TOTAL_YIELD: This is the total yield for the inverter till that point in time. The weather data includes the following information DATE_TIME and PLANT_ID are identical with the description above. Other than that: SOURCE_KEY: Stands for the sensor panel id. This will be common for the entire file because there's only one sensor panel for the plant. AMBIENT_TEMPERATURE: This is the ambient temperature at the plant. The unit for this data is C MODULE_TEMPERATURE: There's a module (solar panel) attached to the sensor panel. This is the temperature reading for that module. Note: After comparing this data with other publications, I assume the correct unit for this data is C IRRADATION: Amount of irradiation for the 15 minute interval. Note: After comparing this data with other publications, I assume the correct unit for this data is kw/m2 III. EXPERIMENT AND ANALYSIS Initially to better understand about the nature and characteristics of the collected data the correlation between different variables are estimated. The Fig. 1 presents the heat map of the correlation analysis between power generated and temperature data available in the dataset. Fig. 1.Heat map of correlation between Power generated and Weather Data The aim of the correlation analysis is to estimate the linear correlation between the variables which helps in removing any redundant or irrelevant information available in the dataset. The correlation can be estimated by www.turkjphysiotherrehabil.org 2502

finding the ratio between covariance of two variables and the product of standard deviation of the two. It can be mathematically expressed as follows ρ X,Y = cov(x, Y) σ X, σ Y Where ρ is the pearson correlation coefficient, and σ is the standard deviation. From the heat map show above it can be inferred that there is high correlation between the ambient temperature and the power generation. From the collected it was noted that few inverters have not received DC power even though the corresponding irradiation was high enough for power generation. This indicates the malfunction or anomaly event present in the data. Further to study this in detail the daily distribution of the power generation and the measured irradiation is plotted as graph in Fig.2& 3 below. The daily distribution plot of DC power generation showed multiple occasions where the DC power generated was zero during day time. The distribution plot of solar irradiation exhibits that the solar radiation never dropped to a lower value at day time. Fig. 2.Daily Distribution of Solar Irradiation Fig.3 Daily Distribution of DC Power Generated The ambient temperature (shown in Fig. 4) recorded at the solar plant and the module temperature measured at the solar panel has been analyzed to understand their range and distribution. The value of ambient temperature ranges between 20 and 35 degree Celsius. The module temperature was observed to higher than the ambient temperature during midday and it ranges between 18 and 65 degree Celsius. www.turkjphysiotherrehabil.org 2503

Fig. 4 Daily and Ambient Temperature on a specific date From the exploratory analysis it was inferred that there is a linear relation between the DC power generated and the solar irradiation. This relation can be modeled using a simple linear relationship as expressed below, P(t) = a + b. E(t) where p(t) denotes the power generated and the E(t) denotes the solar irradiation. The module temperature at few days reached up to 55 degrees but the efficiency of the solar panel is lower at higher ambient temperature. Hence a linear model cannot uncover the non-linear relationship between them. The Fig. 5 below presents the prediction results of a linear regression model for a normal and faulty day and Fig. 6 the residual values are presented for the same. It was observed that on a faulty day the predictions are more erroneous when compared to a normal day. Fig. 5 Plot of Results of Linear regression for a normal and faulty day Fig. 6 Analysis of the Residual values The primary aim of the experiments is to fit the observed data including the temperature and solar irradiation to the generated DC Power. Support vector regression (SVR) is due to the non-linear relation between the given input and output variable. As per the SVR the error in predicting the dependent variable e will be estimated which determines the number of support vectors and when the value of e is smaller, then it indicates a lower www.turkjphysiotherrehabil.org 2504

tolerance state for error. The best fit or the optimal regression line can be estimated based on the following expression. (mx + c) y e andy (mx + c) e When the prediction error is less than the value of e then it is tolerable and hence the data points contributing an error outside the error region e are used for estimation of the final loss in prediction. The support vector training relies only on a subset of the data points and during cost estimation the training data points whose respective error close to e can be ignored. SVR suits best for exploring the non-linear or complex relationship between the dependent and independent variables since the non-linear mapping problem is converted in to estimation of hyperplane in a higher dimensional space. Radial Basis Function (RBF) kernel is used in our experiments and it can be expressed mathematically as below; K(x, x ) = exp( γ x x 2 ) Where γ = 1 2σ 2 and x x 2 denotes the squared Euclidean distance between the feature vectors. Fig. 6 Plot of Results of Linear regression for a normal and faulty day Fig. 7 Plot of Results of SVR Non-Linear regression for a normal and faulty day For analyzing the performance of the linear and non-linear models the residuals of their predictions are plotted in Fig. 8 The performance of the non-linear model Support Vector Regression was observed to be superior when the solar irradiation was higher. www.turkjphysiotherrehabil.org 2505

Fig. 8 Comparison of Linear and NonLinear Model IV. CONCLUSION Based on the non-linear regression model the events such as equipment failure or poor performance of equipment can be identified. The non-linear model was trained on the temperature and the solar irradiation data measured using weather sensors. This approach suits for real time monitoring of the solar plant fault identification due to its accuracy and easier implementation. The proposed approach depends only on available handson information for the fault identification or predicting the future power. Based on the forecasted weather data the near future power generation can be estimated. The condition is that the accuracy of the prediction depends on the accuracy of the forecasted weather data (solar irradiation, and temperature). REFERENCES 1. M. N. Akram and S. Lotfifard, "Modeling and Health Monitoring of DC Side of Photovoltaic Array," in IEEE Transactions on Sustainable Energy, vol. 6, no. 4, pp. 1245-1253, Oct. 2015. 2. Chine W, Mellit A, Lughi V, Malek A, Sulligoi G, Pavan AM, A novel fault diagnosis technique for photovoltaic systems based on artificial neural networks, Renew Energy, 2016. 3. Zhicong Chen, Lijun Wu, Shuying Chen, Peijie Lin, Yue Wu, Wencheng Lin, Intelligent fault diagnosis of photovoltaic arrays based on optimized kernel extreme learning machine and I-V characteristics, in Applied Energy, Elsevier, May 2017. 4. Mekki, H., Mellit, A., Salhi, H., Artificial neural network-based modelling and fault detection of partial shaded photovoltaic modules. Simul.Model.Pract.Theory 67, 1 3., 2016. 5. Y. Zhao, R. Ball, J. Mosesian, J.-F. de Palma, and B. Lehman, Graphbased semi-supervised learning for fault detection and classification in solar photovoltaic arrays, IEEE Transactions on Power Electronics, vol. 30, no. 5, pp. 2848 2858, 2015. 6. S. Rao, S. Katoch, P. Turaga, A. Spanias, C. Tepedelenlioglu, R. Ayyanar, H. Braun, J. Lee, U. Shanthamallu, M. Banavar, and D. Srinivasan, A Cyber-Physical System Approach for Photovoltaic Array Monitoring and Control, in Proc. IEEE IISA, Larnaca, Cyprus, 2017. 7. S. Katoch, G. Muniraju, S. Rao, A. Spanias, P. Turaga, C. Tepedelenlioglu, M. Banavar, D. Srinivasan, Shading Prediction, Fault Detection, and Consensus Estimation for Solar Array Control, 1st IEEE International Conference on Industrial Cyber-Physical Systems (ICPS2018), Saint. Petersburg, Russia, May 2018. 8. G. Muniraju, S. Rao, S. Katoch, A. Spanias, P. Turaga, C. Tepedelenlioglu, M. Banavar, D. Srinivasan, A Cyber-Physical Photovoltaic Array Monitoring and Control System, International Journal of Monitoring and Surveillance Technologies Research, vol., issue 3, May 2018. 9. Fazai, R., et al. "Machine learning-based statistical testing hypothesis for fault detection in photovoltaic systems." Solar Energy 190 (2019): 405-413. 10. Belaout, Abdesslam, FatehKrim, and Adel Mellit. "Neuro-fuzzy classifier for fault detection and classification in photovoltaic module." 2016 8th International Conference on Modelling, Identification and Control (ICMIC). IEEE, 2016. 11. Harrou, Fouzi, et al. "Reliable fault detection and diagnosis of photovoltaic systems based on statistical monitoring approaches." Renewable energy 116 (2018): 22-37. 12. Rao, Sunil, Andreas Spanias, and CihanTepedelenlioglu. "Solar array fault detection using neural networks." 2019 IEEE International Conference on Industrial Cyber Physical Systems (ICPS).IEEE, 2019. 13. Xie, Tuo, et al. "A hybrid forecasting method for solar output power based on variational mode decomposition, deep belief networks and autoregressive moving average." Applied Sciences 8.10 (2018): 1901. 14. Boussaada, Zina, et al. "A nonlinear autoregressive exogenous (NARX) neural network model for the prediction of the daily direct solar radiation." Energies 11.3 (2018): 620. 15. Kumar, M. S., & Neelima, P. (2011). Design and implementation of scalable, fully distributed web crawler for a web search engine. International Journal of Computer Applications, 15(7), 8-13. 16. Balaji, K. "Load balancing in Cloud Computing: Issues and Challenges." Turkish Journal of Computer and Mathematics Education (TURCOMAT) 12.2 (2021): 3077-3084. 17. Sushama, C., Kumar, M. S., & Neelima, P. (2021). Privacy and security issues in the future: A social media. Materials Today: Proceedings. 18. Natarajan, V. A., Kumar, M. S., Patan, R., Kallam, S., & Mohamed, M. Y. N. (2020, September). Segmentation of Nuclei in Histopathology images using Fully Convolutional Deep Neural Architecture. In 2020 International Conference on Computing and Information Technology (ICCIT-1441) (pp. 1-7). IEEE. 19. VA Natarajan, MM Babitha, MS Kumar, Detection of disease in tomato plant using Deep Learning Techniques, International Journal of Modern Agriculture 9 (4), 525-540. 2021. www.turkjphysiotherrehabil.org 2506