2025 International Conference on Sensing, Measurement & Data Analytics in the era of Artificial Intelligence (ICSMD 2025)

Water supply flow forecasting is crucial for urban water management and pipeline optimization. However, existing methods often fail to handle complex temporal patterns and structural components effectively, limiting prediction accuracy. To address these issues, we propose a large time series model(LTSM) with dynamic boundary quantization and Seasonal-Trend decomposition using Loess(STL) to enhance modeling precision and structural adaptability. Specifically, the input time series is first normalized via mean-scaling and decomposed into trend, seasonal, and residual components using STL. Each component is then discretized using a dynamic quantization strategy guided by local statistical features , and the resulting token sequences are concatenated and fed into the LTSM. During inference, the predicted token sequences are decoded, dequantized, and recombined to form the final forecast.Experiments on water flow data demonstrate that our method outperforms SOTA models in terms of MAE, MSE, and R². 

dynamic boundary quantization,flow forecasting,large time series model,Seasonal-Trend decomposition using Loess