Abstract
The Great Lakes are vital freshwater resource for both the United States and Canada. Therefore, the importance of this research lies in its potential to provide timely and accurate information for decision-makers. Improved water level predictions can aid in flood risk management, optimize water resource allocation, and support ecological conservation efforts. This study was conducted to completely eliminate traditional machine learning models’ lag effects with phase space reconstruction (PSR). The prediction was conducted using historical monthly mean water level datasets of Lake Ontario for the period 1918–2023, divided into training (1918–2002) and testing (2003–2023) datasets. The results revealed that PSR- RF outperform the standard random forest, KNN and LSTM models across all metrics, including Correlation Coefficient (0.999), Nash–Sutcliffe Efficiency (0.998), Root Mean Squared Error (0.014), Coefficient of Determination (0.998), and the slope and intercept of the regression equation (𝑦 = 0.98𝑥+1.484).