Pollution index of waterfowl farm assessment and prediction based on temporal convoluted network

Abstract

Environmental quality is a major factor that directly impacts waterfowl productivity. Accurate prediction of pollution index (PI) is the key to improving environmental management and pollution control. This study applied a new neural network model called temporal convolutional network and a denoising algorithm called wavelet transform (WT) for predicting future 12-, 24-, and 48-hour PI values at a waterfowl farm in Shanwei, China. The temporal convoluted network (TCN) model performance was compared with that of recurrent architectures with the same capacity, long-short time memory neural network (LSTM), and gated recurrent unit (GRU). Denoised environmental data, including ammonia, temperature, relative humidity, carbon dioxide (CO2), and total suspended particles (TSP), were used to construct the forecasting model. The simulation results showed that the TCN model in general produced a more precise PI prediction and provided the highest prediction accuracy for all phases (MAE = 0.0842, 0.0859, and 0.1115; RMSE = 0.0154, 0.0167, and 0.0273; R2 = 0.9789, 0.9791, and 0.9635). The PI assessment prediction model based on TCN exhibited the best prediction accuracy and general performance compared with other parallel forecasting models and is a suitable and useful tool for predicting PI in waterfowl farms.