Reservoir Inflow Prediction Using Time Lagged Recurrent Neural Networks

Although there have been many successful applications of artificial neural networks (ANNs) to capture non linear relationship of inflow into a reservoir, there are cases when ANNs have not been able to predict flow extremes accurately. Applicability of non linear model based on ANN with a random component embedded is explored for Pawana reservoir in Upper Bhima River Basin, Maharashtra, India. Suitability of time lagged recurrent networks (TLRNs) with time delay, gamma and laguarre memory structures is investigated for predicting seasonal (June to October) reservoir inflow with a monthly time step. Trial and error procedure is adopted in selecting input nodes and hidden nodes. Input to the network includes preceding records of reservoir inflow (varying from one lag to four lags). Performance of back propagation through time (BPTT) algorithm is investigated for various inputs assuming different training and testing combinations. Due to large variation in the observed data values, transformation of the series is also tried and found that the network trained for 50:50 training and testing data sets predicted better values. The validation of the models was performed using comparison of principal statistics (mean, standard deviation, skewness and kurtosis), goodness-of-fit measures (MSE, MAE, MRE and R), time series plots and scatter plots. Encouraging results obtained in the present study indicated that the log-transformed, BPTT trained TLRN resulted in better and reliable forecast of extreme high and low inflows as compared to non-transformed series.