Anomaly Detection And Short Term Forecasting

Abstract

Forecasting of Electric load always had an empirical impact on power operations. It is the most challenging and critical attribute of energy management system planning and operation. With the emergence of smart grid (SG) technology, the conventional grid has been promoted to a modern digital grid which ensures robustness, reliability, safety, and economic viability to the existing power systems. In contrast, it has introduced complex information and greater threats for electric load forecasting. As a result, cybersecurity has become a major challenge for the existing grid. Cyber attacks have numerous types but the most important one is the data integrity attack which results to reflect wrong projections. The data integrity attacks are termed as anomalies in time series data. Such anomalies can malignantly reshape a specific segment of input load data, which results in a remarkable deviation from original values. In this thesis, an anomaly detection model and a short-term load forecasting based model have been developed. Initially, anomaly detection has been implemented on electrical load for the data of duration three years. The LSTM model is trained and used to predict the electrical load and the mean square error is computed on test data. Based on maximum observed error, a threshold is selected to detect anomalies in the total data. After anomaly detection, the short-term load forecasting is performed using LSTM and Bi-LSTM methods and the results are compared on the basis of Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE). Additionally, the computational complexity analysis for both algorithms is carried out. For short-term forecasting, hourly data of electrical load is considered. The Bi-LSTM model produced the MAPE of 8.79% while simple LSTM model showed 14.85%. The Bi-LSTM outperformed the conventional LSTM model with a margin of 6% MAPE overall. However, the computational time of LSTM model was short, almost 60% of the time taken by the Bi-LSTM model.