Bayesian Optimization-based Multilayer Perceptron Prediction Algorithm for Catfish Feeding Quantity

Abstract

At present, the cost of aquaculture is continuously rising, with feed costs accounting for a significant proportion. Most feeding methods rely on empirical judgment and lack long-term tracking and prediction capabilities, which can easily lead to feed waste, water pollution, and negative impacts on the growth of aquatic animals. To address the lack of long-term prediction ability, the main factors affecting catfish feeding were first analyzed, water temperature, dissolved oxygen content, ammonia nitrogen content, average weight of fish, the total number, and meteorological conditions were selected as the input variables for the feeding calculation model, a multilayer perceptron (MLP) feeding calculation model was established, the data was normalized, and a model training dataset was constructed. After the model was established, a Bayesian optimization algorithm was used to optimize the hyperparameters of the MLP model to enhance its performance. Finally, a detailed comparative analysis between the actual feeding quantity and the model-predicted feeding quantity, using data from the three breeding ponds in the test set, verified the model’s performance and generalization ability. The average absolute percentage error remained below 4%, and the absolute error was controlled at less than 700kg during August and September, in the peak feeding period for the 11# ponds. However, during special periods such as fish disease, harvesting, and fishing in the breeding ponds, breeders would adjust the feeding amount accordingly, leading to a large error between the model-predicted feeding amount and the actual feeding amount. This also highlighted the model’s limitations in coping with extraordinary events. Meanwhile, the traditional MLP model, random forest model, support vector machine model, and BO-MLP model were selected for comparative tests to verify the superior performance of the described model. Overall, the research results presented were of great significance as an important decision support tool for the aquaculture industry, improving feed utilization and enhancing aquaculture efficiency.

Keywords:aquaculture, Bayesian optimization, multilayer perceptron, feeding quantity, prediction algorithm