Time-series foundation modeling enables accurate lake ecosystem forecasting

Abstract

Accurate ecological forecasting is increasingly essential for understanding and managing ecosystem responses to climate variability and anthropogenic pressures; however, prediction remains difficult in lakes because key biological variables, such as phytoplankton biomass, exhibit nonlinear dynamics, observational noise, data sparsity, and non-stationarity. This study aimed to test whether large pre-trained time-series foundation models can improve forecasts of lake phytoplankton dynamics, focusing on chlorophyll-a (Chl.a) and water-quality variables in two ecologically contrasting Japanese lakes, the deep-stratified Lake Biwa and the shallow nutrient-rich Lake Kasumigaura. Using extensive monthly monitoring records spanning up to 30 years, we benchmarked two time-series foundation models—Transformer-based Chronos-T5 and probabilistic Lag-Llama—against 15 statistical (AR, ARIMA, SARIMA, Prophet), machine learning (Random Forest, XGBoost, KNN, SVR), and deep learning (LSTM, CNN, TCN, and SSA-hybrid) approaches. Chronos-T5 achieved superior reproduction accuracy and stability across diverse environmental variables and outperformed all other tested models. This performance stemmed from its ability to capture long-term dependencies and complex temporal patterns through large-scale pre-training. We further identified an optimal training window of approximately 14 years, balancing data sufficiency with regime stability, beyond which model accuracy diminished owing to ecological regime shifts. Our findings highlight the transformative potential of time-series foundation models for ecological forecasting, providing a scalable, data-driven framework that can underpin robust early warning systems and adaptive management strategies for aquatic ecosystems worldwide. Significance statement Accurate forecasting of lake ecosystems is crucial for water management under climate change, yet predicting phytoplankton dynamics remains challenging due to complex, non-stationary biological responses. Here, we demonstrate that large time-series foundation models significantly outperform traditional statistical, machine learning, and deep learning methods in forecasting chlorophyll-a and water-quality dynamics across contrasting Japanese lakes. Pre-trained on vast temporal datasets, these models successfully capture long-term dependencies and non-linear patterns without domain-specific architecture tuning. Furthermore, we identified an optimal 14-year training window, highlighting how ecological regime shifts constrain model memory. Our findings introduce time-series foundation models as a scalable, high-precision paradigm for ecological forecasting, offering vital tools for global freshwater conservation and early-warning systems.