Improving time series forecasting using deep learning

The dissertation’s primary goal is divided into two parts. The first part of the dissertation focuses on developing a time series model using the multiplicative decomposition approach to break down time-series data into trend-cycle-irregular and seasonality components. After that, Seasonal Autoregressive Integrated Moving Average with Exogenous Variables (SARIMAX) is used to select the model that best captures the trend-cycle-irregular portion. The SARIMAX model’s trend-cycle-irregular element is then coupled with the seasonal index to provide a series of forecast values as a linear component model.To determine the residuals from the first part, the second part uses the first part as a linear component model. Deep learning, such as Artificial Neural Network (ANN) and Long Short-Term Memory (LSTM) as a nonlinear component model, represents the second half. Finally, we obtain a novel hybrid model with improved performance by utilizing deep learning as a linear additive combination model. The suggested forecasting approach is tested on six different types of actual monthly datasets, including the electricity consumption in provincial area of Thailand, the diesel consumption of Thailand, the temperature in Bangkok, Thailand, the USD/THB exchange rate, the SET index, and the JPY/THB exchange rate. We employed the four-performance metrics, namely, Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and Reduce Error Rate, to evaluate hybrid model prediction accuracy and to compare multiple models fitted to a time series (RER). We have provided the produced forecast diagram for each of the six datasets, which visually displays the similarity between the original and forecasted observations.The Experimental result shows that the new hybrid forecasting using decomposition with the SARIMAX model and LSTM (DEC-SARIMAX-LSTM) can perform well. The best hybrid model has a reduced average error rate for three months, 12 months, and 24 months lead time forecasting of 93.7404%, 92.5161%, and 90.2869%, respectively. In addition, the new hybrid forecasting model between the decomposition method with the SARIMAX model and LSTM (DEC-SARIMAX-LSTM) has the lowest average MAPE of 0.1450% for three months, 0.2339% for 12 months, and 0.3296% for 24 months lead time forecasting, respectively. The best forecasting model has been checked by using residual analysis. We conclude that the combined model is an effective way to improve more accurate forecasting than a single forecasting method.