Adaptive Holt Winter Autoregressive Model for IoT Forecasting

Abstract

In business and industry, forecasting the future values of a time series is an age-old and widely utilized data analysis strategy. IoT temporal data, used in various applications, is one of the most well-known examples of such a time series. The main problem is to forecast a future window of the data using the provided IoT temporal data. Many forecasting models that deal with such data were proposed, such as Rolling Window, SVR-RBF, ARIMA, etc. However, in the training process, these models use all available data or a previous direct window of data. As a result, the training data may contain irrelevant data patterns to the current situation, lowering the total forecasting accuracy. This paper proposes a new adaptive forecasting model called AHW-SWA by creating two models based on the Exponential Smoothing method. The first model is trained using the previous direct window of data. The second model uses an extracted data window from the far historical data with a pattern most similar to the current situation. Then an adaptive process is used to balance the predicted values from both models to achieve higher accuracy. AHW-SWA model reduces the overall forecasting error by comparing the extracted and the latest data patterns. Our model outperforms well-known models, including Rolling Window, SVR-RBF, and ARIMA.