Time-series foundation modeling enables accurate lake ecosystem forecasting

Fine-tuned Chronos-T5 secured Rank 1 across all six monitoring sites, achieving R2 > 0.80-1.00 for dissolved oxygen, water temperature, pH, and nutrients (NH4, NO3, PO4). Prior ecological forecasting relied mainly on linear statistical models such as ARIMA, machine learning methods such as random forest and XGBoost, and deep learning architectures such as LSTM and CNN; this study is the first to systematically bring large pre-trained time-series foundation models into lake water-quality forecasting with a cross-model benchmark. Based on long-term monitoring records from two sites in Lake Biwa (monthly data from January 1994 to December 2023, 30 years) and four sites in Lake Kasumigaura (monthly data from January 1997 to December 2012, 16 years), compared against 15 alternative models under a common R2 metric, with 2023 (Lake Biwa) and 2012 (Lake Kasumigaura) used as independent holdout validation years.

Within the foundation-model category, Chronos-T5 excelled while Lag-Llama showed severe degradation (R2 approximately 0), indicating that domain adaptation and pre-training architecture critically determine foundation-model efficacy in aquatic time-series forecasting. This contrast shows that foundation models are not a homogeneous class: differences in architecture, parameter scale, and training methodology substantially change ecological time-series performance, with the text noting Chronos-T5 at approximately 710MB versus Lag-Llama at approximately 2.1MB. A direct comparison of two foundation models on the same data and evaluation pipeline, where Lag-Llama generated probabilistic forecasts by sampling 100 trajectories and averaging results, while Chronos-T5 was fine-tuned from the pre-trained amazon/chronos-T5-large model.

Chlorophyll-a was the most challenging variable, with R2 below 0.30 across all models including Chronos-T5, whereas physical and stable chemical parameters such as water temperature and dissolved oxygen showed high baseline predictability across most models. The result identifies parameter-intrinsic complexity, rather than model choice, as the primary bottleneck, noting that chlorophyll-a is driven by highly nonlinear, episodic algal dynamics that require specialized modeling strategies. Based on R2 comparisons across 17 models (including two foundation models) at six sites for chlorophyll-a and each water-quality parameter, with classical machine learning models (KNN, XGBoost, Random Forest) and statistical methods (Prophet) occasionally achieving top ranks at specific localized sites such as NH4 at Imazu or Minamihira.

An optimal training window of approximately 14 years was identified for both lakes, with model performance showing a nonlinear relationship to training data length, a local peak at 3-4 years, and degradation when training extended beyond 15 years. The finding brings ecological regime-shift theory into the choice of training data length, suggesting that data spanning regime shifts may introduce noise, and offering guidance for model retraining and data selection strategies. Training windows were varied from 1 to 29 years for Lake Biwa and 1 to 15 years for Lake Kasumigaura, with peak normalized values used to identify the optimal training period, and with SSA smoothing observed to reduce forecasting accuracy, indicating that ecological noise contains meaningful biological signal.

The study targets lake systems with long-term monthly continuous monitoring records (at least 10 years of chlorophyll-a data) and applies to water-quality forecasting scenarios where chlorophyll-a, water temperature, dissolved oxygen, pH, and nitrogen and phosphorus nutrients are the prediction targets, with conclusions grounded in six sites across two Japanese lakes, Lake Biwa and Lake Kasumigaura. Methodologically, Chronos-T5 was fine-tuned with a 12-month forecast horizon, training windows ranged from 1 to 29 years for Lake Biwa and 1 to 15 years for Lake Kasumigaura, and validation used the holdout years 2023 and 2012. The framework offers a starting point for combining foundation models with mechanistic ecological understanding and climate projections, and for developing visualization pipelines aimed at managers.

A careful reader might watch whether the chlorophyll-a bottleneck, with R2 below 0.30 across all models, eases when mechanistic information or higher-frequency observations are introduced; whether the approximately 14-year optimal training window holds in other lakes with different regime-shift histories; how much of Lag-Llama's degradation relates to model size, pre-training data, or domain adaptation; and how probabilistic forecast calibration and uncertainty characterization perform beyond R2 as a single metric. In addition, this is a preprint that has not yet been peer reviewed, and figure details require reading the original figures.