Comparison between traditional methods and artificial neural networks for ammonia concentration forecasting in an eel (Anguilla anguilla L.) intensive rearing system

Abstract

One of the main problems in the management of fishfarms with water recirculating system is the forecasting and control of ammonia concentration in order to minimise the fish stress status. This paper examines methodologies of prediction in a real-time environment for an eel intensive rearing system. Approaches based on linear multiple regression, univariate time series models (exponential smoothing and autoregressive integrated moving average (ARIMA) models) and computational neural networks (ANNs) are developed to predict the daily average ammonia concentration in rearing tanks with water recirculating. The models are established using actual data from an eel fishfarm in southern Spain. The input variables used in the models (multiple regression, Holt smoothing model, ARIMA models and ANN models) are the ammonia concentration of previous days. In ANN models, the training method used is a standard back-propagation variation known as extended-delta-bar-delta (EDBD). Different neural architectures, whose learning is carried out by crossvalidation and controlling several threshold determination coefficients, are compared. Globally, the nonlinear ANN model approach is shown to provide a better prediction of daily average ammonia concentration than linear multiple regression and univariate time series analysis when the correlation between data series is low and when the models were obligated to predict in a situation for which specifically had not been calibrated. The best results were obtained by 5:10s:15s:1l ANN model in the pre-growth series.