Application of optimized Support Vector Machine in monthly streamflow forecasting: using Autocorrelation Function for input variables estimation

Hydrologic forecasting serves as an important tool in water resource management to mitigate disasters and managing water infrastructures. In field of hydrology, models for time series forecasting have been developed throughout the years, including the application of data-driven models. In this research, application of Artificial Intelligence (AI) technique in Support Vector Machine (SVM) method is used to forecast monthly discharge in Pemali River Basin, Indonesia. SVM method, optimized with fast messy genetic algorithm (fmGA), is employed for time series forecasting, whereas input data rely solely on previous values. Model performance is assessed with three different performance metrics and against Seasonal Autoregressive Moving Average (SARIMA) method for comparison. Scenarios are constructed with different input pattern for SVM to identify appropriate input data for giving good prediction accuracy. Input data are developed with and without selecting mechanism. Selecting mechanism is done based on assessment in Autocorrelation Function (ACF) coefficient of the time series data. While input data without the selecting mechanism consist of monthly discharge up to lag time 12 months prior (Q_t-1, Q_t-2,…,Q_t-12). The result shows that input data with good correlation can give good prediction accuracy. Involvement of poorly correlated input data series may decrease model performance. However, with proper combination of input data, SVM can have good forecasting accuracy regardless of having the poorly correlated input data. Coherently, appropriate input data combination can reduce the number of support vectors in SVM, thus scaling down the risk of over fitting data.