NadamClip: A Novel Optimization Algorithm for Improving Prediction Accuracy and Training Stability

Accurate prediction of key environmental parameters is crucial for intelligent control and optimization, yet it remains challenging due to gradient instability in deep learning models, like Long Short-Term Memory (LSTM), during time series forecasting. This study introduces a novel adaptive optimization algorithm, NadamClip, which integrates gradient clipping directly into the Nadam framework to address the trade-off between convergence efficiency and gradient explosion. NadamClip incorporates an adjustable gradient clipping threshold strategy that permits manual tuning. Through systematic experiments, we identified an optimal threshold range that effectively balances model performance and training stability, dynamically adapting to the evolving convergence characteristics of the network across different training phases. Aquaculture systems are regarded as similar to modern biomanufacturing systems. The study evaluated an aquaculture dataset for ammonia concentration prediction in aquaculture environmental control processes. NadamClip achieved outstanding results on key metrics, including a Root Mean Square Error (RMSE) of 0.2644, a Mean Absolute Error (MAE) of 0.6595, and a Coefficient of Determination (R2) score of 0.9743. Compared to existing optimizer enhancements, NadamClip pioneers the integration of gradient clipping with adaptive momentum estimation, overcoming the traditional paradigm where clipping primarily serves as an external training control rather than an intrinsic algorithmic component. This study provides a practical and reproducible optimization framework for intelligent modeling of dynamic process systems, thereby contributing to the broader advancement of machine learning methods in predictive modeling and optimization for data-driven manufacturing and environmental processes.