Harmful algal blooms are expanding globally across freshwater ecosystems due to climate change and accelerating eutrophication, posing escalating threats to public health and economic stability. Robust early warning requires predictive frameworks that are accurate at hourly resolution, uncertainty-aware, and mechanistically interpretable. Existing process-based models are limited by sparse parameterization, while statistical and deep-learning approaches typically operate at daily resolution, produce only point predictions, and rely on threshold-based classifications or post-hoc explanations that fail to capture high-frequency temporal dynamics or quantify predictive uncertainty, thereby hindering proactive intervention and masking hidden ecological risks. Here we present BloomNet, an intrinsically interpretable deep neural network architecture configured for multi-horizon quantile forecasting, resolves these limitations by integrating future environmental covariates with historical observations from a hyper-eutrophic lake. Evaluated on a four-year hourly record, BloomNet achieves exceptional predictive stability across 24-, 48-, and 72-h horizons (R² up to 0.78, 24-h MAPE = 25.7%), outperforming long short-term memory, Transformer, and temporal convolutional network baselines while circumventing iterative error accumulation. The network's dual-path variable-selection and attention mechanisms decode shifting ecological drivers directly across horizons, revealing a transition from short-term thermal regulation (water temperature weight up to 0.640) to medium-term nutrient governance (total phosphorus weight up to 0.689), while isolating distinct multimodal attention spikes 24h prior to bloom onset. Real-time interpretability further reveals that driver importance and temporal dependencies shift dynamically between bloom-onset and non-bloom states. These probabilistic quantiles establish a graduated, risk-oriented warning protocol that transforms reactive water resource administration into precision ecotechnology.

Escalating global temperatures threaten economic stability by worsening occupational heat stress and reducing workforce productivity. Despite advancements in macroeconomic modeling, current risk assessments rely on coarse annual aggregations and ignore internal labor mobility, thereby masking highly unequal sub-national vulnerabilities and underestimating how labor mobility buffers cascading supply-chain losses. Here we present a high-resolution, agent-based dynamic supply chain network model that integrates empirical daily mobility data across 313 Chinese cities to quantify the spatiotemporal cascading economic impacts of occupational heat exposure. We show that annual heat stress costs China 2933.5 billion CNY (2.6% of GDP), with systemic propagation through supply chains driving 59% of these losses. Crucially, labor mobility redistributes risk: net labor inflows into industrialized, high-heat southeastern regions provide a factor-compensation effect that buffers cascading losses by offsetting direct local productivity shocks, saving a net 7.2 billion CNY directly and 24.6 billion CNY indirectly nationwide. Under a 2030 warming scenario (SSP3-7.0), total losses expand 1.6-fold to 4672.9 billion CNY, though integrated multi-level adaptations—combining industrial restructuring with work-hour shifting—can mitigate these future losses by 30%. These findings reveal that demographic mobility dictates the economic geometry of climate vulnerability, highlighting that resilient climate adaptation requires synchronized network-level interventions rather than isolated local policies.