Fish weight prediction using empirical and data-driven models in aquaculture systems

Predicting fish growth trajectory is crucial in achieving better control for the growing phase of the fish in aquaculture systems. However, fish biomass sensors that directly measure the fish's weight are currently unavailable. Additionally, the fish growth model that relies on the fish weight is affected by environmental and feeding management factors, raising the difficulty of predicting the weight. This paper proposes comparative empirical and data-driven models for fish weight prediction. The empirical fish growth model relies directly on the growth ratio, an essential component for optimizing fish growth, while the data-driven model is based on long short-term memory (LSTM). Since the exact daily food ratio, proportional to the body weight of a growing fish, cannot be prescribed, we optimize the growth ratio of the empirical model in a sliding window framework based on the previous growth periods and then predict the fish weights a few days ahead through a nonlinear exponential regression approach. The LSTM method uses a normal distribution to generate multiple datasets over the experimental fish weight range and interpolate these data to provide a good prediction. To this end, we propose a weighted sum optimization method that combines the individual cost of the LSTM and empirical methods to achieve better short-term prediction performance. The simulation results demonstrate the effectiveness of the proposed models in predicting the fish weight and show that the combined weighted sum approach reduces the testing errors by 61.3% and 70.9% compared to the empirical and LSTM methods, respectively. Furthermore, an increasing specific growth rate (SGR) for both LSTM and combined models was observed in the short-term predictions.