Data Augmentation Technique Based on Improved Time-Series Generative Adversarial Networks for Power Load Forecasting in Recirculating Aquaculture Systems

Factory aquaculture faces a difficult situation due to its high running costs, with one of the main contributing factors being the high energy consumption of aquaculture workshops. Accurately predicting the power load of recirculating aquaculture systems (RAS) is critical to optimizing energy use, reducing energy consumption, and promoting the sustainable development of factory aquaculture. Adequate data can improve the accuracy of the prediction model. However, there are often missing and abnormal data in actual data detection. To solve this problem, this study uses a time-series convolutional network–temporal sequence generation adversarial network (TCN-TimeGAN) to synthesize multivariate RAS data and train a long short-term memory (LSTM) network on the original and generated data to predict future electricity loads. The experimental results show that the data generated based on the improved TCN-TimeGAN provide more comprehensive coverage of the original data distribution, with a lower discriminative score (0.2419) and a lower predictive score (0.0668) than the conventional TimeGAN. Using the generated data for prediction, the R2 reached 0.86, which represents a 19% improvement over the ARIMA model. Meanwhile, compared to LSTM and GRU without data augmentation, the mean absolute error (MAE) was reduced by 1.24 and 1.58, respectively. The model has good prediction performance and generalization ability, which benefits the RAS energy saving, production planning, and the long term sustainability of factory aquaculture.