Abstract
The complementary strengths and weaknesses of both statistical modeling paired with machine learning has been an ongoing technique in the development and implementation of forecasting models that analyze the dataset’s linear as well as nonlinear components in the generation of accurate prediction results. In this paper, autoregressive integrated moving average (ARIMA) and artificial neural networks (ANN) were implemented as a hybrid forecasting model for a power utility’s dataset in order to predict the next day’s electric load consumption. ARIMA and ANN models were serially developed resulting to the findings that out of the twelve evaluated ARIMA models, ARIMA (8,1,2) exhibited the best forecasting performance. After identifying the optimal ANN layers and input neurons, this study showed that out of the six evaluated supervised feedforward ANN models, the ANN model which employed Hyperbolic Tangent activation function and Resilient Propagation training algorithm also exhibited the best forecasting performance. With Zhang’s ARIMA and ANN hybridization technique, this study showed that the hybrid model delivered Mean Absolute Percentage Error (MAPE) of 4.09% which is within the 5% internationally accepted forecasting error for electric load forecasting. Through the findings of this research, both the ARIMA statistical model and ANN machine learning approaches showed promising results in being implemented as a forecasting model pair to analyze the linear as well as non-linear properties of a power utility’s electric load data.
Highlights
Use of individual machine learning and statistical modeling has been in the forefront of predictive analytics due to their promising abilities to deliver close to accurate forecasting results
With Zhang’s Autoregressive Integrated Moving Average (ARIMA) and Artificial Neural Networks (ANN) hybridization technique, this study showed that the hybrid model delivered Mean Absolute Percentage Error (MAPE) of 4.09% which is within the 5% internationally accepted forecasting error for electric load forecasting
This study attempted to present a hybrid model of ARIMA and ANN in load forecasting
Summary
Use of individual machine learning and statistical modeling has been in the forefront of predictive analytics due to their promising abilities to deliver close to accurate forecasting results. Compared to that of ARIMA, ANN has the ability to learn from non-linear datasets due to its strength of being adaptively formed from the implemented features of its own dataset Despite their differences in the kind of data that they can accommodate, ARIMA and ANN hybrid forecasting methodologies as well as modelling techniques are being widely developed due to the potential of generating better predictive performance than individually utilizing each model [2, 6]. For the purpose of optimal predictive performance, the main challenge in the hybridization of these machine learning and statistical modeling approaches relies on the optimal match between the data they are processing along with the forecasting ability that they enforce in their inherent unique advantages This gives data modelers the challenge beyond the functions of data preparation and explore on the performance analysis of the ARIMA and ANN hybridization technique that can yield optimal predictive results
Sign-in/Register to view highlights & summary 
Published Version (
Free)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call