Floods are among the most devastating natural disasters worldwide, inflicting severe damage on human life, infrastructure, and socioeconomics. Long-term flood forecasting is crucial for sustainable flood risk management, necessitating the development of accurate and efficient prediction models. This study aims to demonstrate the application of machine learning models in long-term flood risk prediction on the downstream watershed of the Ottawa River in Ontario, Canada. Specifically, it investigates the use of Support Vector Machine (SVM), Artificial Neural Network (ANN), and Extreme Learning Machine (ELM) models for long-term flood forecasting. An additional objective is to assess the impact of various variables on flood risk assessment, utilizing the Pearson method to compare the correlation of inputs like precipitation, rainfall, snow, temperature, wind speed, and humidity with the output water level index. The performance of the three applied models for flood risk forecasting was evaluated. Results indicate that the ELM model outperforms the others, achieving a higher accuracy in both training and testing phases. The ELM model yielded very high correlation coefficients (R) of 0.860 and 0.901; low Root Mean Square Error (RMSE) values of 0.417 and 0.374; and Mean Absolute Error (MAE) values of 0.523 and 0.463 for the training and testing phases, respectively. In addition, a sensitivity analysis of the best-calibrated ELM model revealed a significant dependency on the humidity parameter. The findings underscore the potential of machine learning models, particularly the ELM, in enhancing long-term flood forecasting accuracy. This research contributes to the growing body of knowledge on machine learning applications in natural disaster risk assessment and offers valuable insights for effective flood risk management.