An Approximation Computation Approach to the Enhancement of Big Data Analysis Methods for Time-Series Anomaly Detection of PV Systems

This paper follows the principles of the Big Data Analysis (BDA) framework to construct a set of experiments using a photovoltaic (PV) system dataset, providing an empirical evaluation of these principles. We propose a novel two-dimensional approach to improving time-series clustering and anomaly detection. Firstly, we enhance the DBSCAN clustering algorithm with LSTM-based residual forecasting (LSTM+DBSCAN) to capture temporal dependencies. Secondly, we incorporate incremental data ingestion to assess the scalability and adaptability of the framework. Our methodology builds upon the Approximation Computation Approach (ACA), which emphasizes a Pattern-Method-Assertion (PMA) framework for data analysis. By leveraging LSTM to extract meaningful temporal patterns and DBSCAN for clustering, we validate and extend the principles of ACA to the context of time-series anomaly detection. Our results demonstrate that the framework could provide guidelines for experiments in both the design of the methods and the datasets.