This study addresses the challenge of estimating Biochemical Oxygen Demand over 5 days (BOD₅) in a remote polar lake, where direct measurement is expensive, time-consuming, and often constrained by limited monitoring data. Using more than four decades of historical water quality records from Lake Oulujärvi, Finland, the research developed predictive models based on observations collected from 34 stations across six sub-basins, with particular attention to stratified periods and depth-resolved measurements. The data were first preprocessed through normalization, detection of missing values, removal of invalid observations, and stratification analysis to improve consistency and model reliability. Three machine learning approaches, including Convolutional Neural Networks (CNN), Random Forest (RF), and XGBoost, were then trained to estimate BOD₅ from routinely monitored water quality variables. The selected predictors included dissolved oxygen (DO), pH, turbidity, electrical conductivity, Secchi depth, water temperature, and sampling time, all of which are readily measurable and physically linked to BOD₅ dynamics. Model performance was optimized using grid search and cross-validation. Among the tested models, the CNN achieved the best results, with validation R² values of 0.88–0.89, followed by RF, and XGBoost. Feature importance analysis highlighted DO, pH, turbidity, electrical conductivity, and sampling time as the most influential variables. The findings show that machine learning can provide accurate, cost-effective, and scalable alternatives for BOD₅ estimation in low-accessibility aquatic systems, supporting improved water quality assessment and management in fragile lake ecosystems under environmental change.