Interpreting the Effect of Generative Adversarial Network Application on Deep Learning Model Performance for Chlorophyll‐ <i>a</i> Concentration Prediction in a Stream Using Explainable Artificial Intelligence

Predicting algal blooms is crucial for effective water quality management. Recent studies have leveraged advanced data-driven models, such as deep learning, for this purpose. However, the effectiveness of these models heavily relies on the availability of high-quality data, which is often costly and time-intensive to collect in real-world environments. This study employed a time-series generative adversarial network (GAN), a representative generative artificial intelligence (AI) model, to produce synthetic data and evaluate its impact on the performance of a long short-term memory (LSTM) network, a widely used deep learning model for time-series prediction. The input variables were constructed with sequence lengths of 3, 6, 9, 12, 15, and 18. Two modeling scenarios were analyzed: one using only real data (LSTM_REAL) and another (LSTM_GAN) that applies knowledge distillation to incorporate information learned from both real data and GAN-generated data. Among the two scenarios, LSTM_GAN with a sequence length of 6 achieved the best performance with an NSE of 0.802, and the results indicated that the impact on performance varied depending on the sequence length. LSTM_GAN showed slight improvement over LSTM_REAL at sequence lengths 3, 6, 9, and 12, while showing degradation at 15 and 18, and the overall effect on performance was modest. However, a quantitative assessment using Shapley value analysis, a well-known explainable AI technique, revealed that the GAN-generated data accounted for 14.5%-24.3% of the total variable importance depending on the sequence length. These findings indicate that GAN-generated data meaningfully influence the model's internal inference process. Overall, this study demonstrates the potential of the GAN algorithm in improving algal bloom prediction models.