ABSTRACT
Rathnapura is a City in Sabaragamuwa Province which is frequently affected by floods due to the over flow of Kalu Ganga. This research was carried out to study the ability to use machine learning techniques to predict and forecast the water level of Kalu Ganga and hence forecast flood events. Water level data for the river Kalu Ganga and its six sub rivers were obtained from the remote sensors deployed by the Irrigation Department. In the research, uni-variate analysis was done using a data set of water level of Kalu Ganga from January 2020 to January 2021 which has a sample time of 1 hour. Here, the lags of Kalu Ganga was used in modelling the water level as a time series. Optimal number of lags was discovered using AIC (Akaike information criterion). Accordingly, ARIMA(2,1,2) model was found to be the statistically best ARIMA model. In multi-variate analysis, a data set of Kalu Ganga and six sub rivers collected from 7th to 28th of April 2021 which has sample time of 1 minute was used. From six sub rivers, two of them join the main river before Rathnapura town and among them Wey Ganga has the highest correlation with water level at Rathnapura. Other four sub rivers also affect the water level at Rathnapura due to the backward effect. A linear regression model and a neural network were created using the above parameters. Performance comparison showed that neural network model perform better due to its capability of handling non-linearities.
Key words: Kalu Ganga, flood, machine learning, time series, linear regression,neural networks, correlation