Iterative Multi-Task Learning on Spatial Time Series Data with Applications to Improvement of Performance Prediction and Monitoring for Solar Panels

Health condition monitoring and failure detection play a crucial role in estimating the performance of solar panels such as degradation trend over time and occurrence of failures. Monitoring and detecting significant degradation can help solar panel owners establish as-needed maintenance strategies on a timely manner. But in some occasions, degradation trend estimation becomes challenging due to limited availability of training data such as many missing observations in time series over a large time span and a lack of history of failure records that are sufficient to establish statistical models. To fill the gap, this thesis proposes a new approach of iterative multi-task learning of Gaussian process in time series data (MTL-GP-TS) by sharing information among similar-but-not-identical datasets from multiple solar panel locations. The proposed MTL-GP-TS model learns unobserved or missing values in a particular time series dataset to forecast the future trend with autoregressive integrated moving average (ARIMA) model, resulting in substantial improvement of forecast over conventional time series modeling approaches. Moreover, the estimated degradation trend with proposed MTL-GP-TS method has the potential to improve the monitoring of significant performance degradation compared with the conventional time series model. This thesis also studies the effect of temporal dependent weather factors on the solar panel performance by integrating a covariate with the MTL-GP-TS algorithm. A case study has demonstrated that inclusion of weather factors into the monitoring of degradation with PV-Weather data integration model can significantly improve the solar panel performance prediction.