Comparative Evaluation of Hybrid SARIMA and Machine Learning Techniques Based on Time Varying and Decomposition of Precipitation Time Series

Accurate precipitation forecasts are much attractive due to their complexity. This study aimed to use the hybrid Seasonal Autoregressive Integrated Moving Average (SARIMA) model and machine learning techniques such as Artificial Neural Networks (ANN) and Support Vector Machines (SVM) to improve precipitation forecasts. Time variation analysis and time series decomposition were the two concepts applied to construct the hybrid models. The performance of the two concepts was evaluated with monthly precipitation time series of two stations in northern Iran. Time variation analysis of time series was conducted with the clustering analysis, which increased the accuracy of forecasting with 20.99% decrease in the geometric mean error ratio for the two stations. SVM model decreased the forecasted error compared to ANN in the internal process of time variation analysis. Average of Mean Relative Error (MRE) were MRESVM= 0.72, MREANN= 0.89, and Mean Absolute Error (MAE) in the two stations were MAESVM= 18.02 and MAEANN= 23.88. Therefore, SVM outperformed the ANN model. Comparison of the two hybrid models indicated that more accurate results belonged to the concept of time series decomposition (the decrease in root mean square error from time variation to time series decomposition concepts was 13.35%). Extracting the pattern of data with SARIMA-based hybrid model with time series decomposition improved the precipitation forecasting. Configurations related to nonlinear components of time series with time steps of residual had good performance (the average of agreement index was 0.9). The results suggest that the hybrid model can be a valuable and effective tool for decision processes, and time series decomposition to linear and nonlinear components has a better performance.

Comparative Evaluation of Hybrid SARIMA and Machine Learning Techniques Based on Time Varying and Decomposition of Precipitation Time Series Keywords:

Support Vector Machines , Cluster analysis , Nonlinear component , Configuration