基于物理概化与深度学习耦合的葛洲坝水位预测模型
A Water-Level Prediction Model for Gezhouba Based on the Coupling of Physical Conceptualization and Deep Learning

针对葛洲坝等低水头、小库容径流式水电站日内发电调度中水位预测精度不足的问题，构建了物理概化模型与长短期记忆网络(LSTM)耦合的水位预测方法。河段传播采用时移与幅度修正概化，历史水量平衡用于反算耗水率样本，LSTM学习耗水率随运行状态变化的时序关系。预测时结合上游下泄、本站计划出力及泄流，按水量平衡递推计算坝前水位。依据2025年枯水期调峰出力线型，选取6个小幅调峰且日内出力计划调整幅度较小的典型日工况，与直接水位LSTM进行比较。耦合模型在4个工况中同时降低平均绝对误差(MAE)和均方根误差(RMSE)，6个工况的平均MAE和RMSE分别由0.13 m、0.15 m降至0.07 m、0.08 m；平均日末绝对误差由0.27 m降至0.03 m。在小幅调峰、运行工况相对稳定的条件下，该方法可为径流式水电站日发电计划编制提供依据。
Water-level prediction accuracy remains insufficient for intraday generation scheduling at low-head run-of-river hydropower stations with limited storage capacity, such as Gezhouba. The proposed method couples a physics-based conceptual model with a long short-term memory (LSTM) network to predict water levels. A time shift and an amplitude adjustment represent flow propagation along the river reach. The method back-calculates samples of the generation water-consumption rate from historical water balances, and the LSTM learns how this rate varies with operating conditions over time. During prediction, water-balance calculations use upstream releases, the plant’s scheduled power output, and spill discharge to recursively determine the forebay water level. The evaluation compares the coupled model with a direct water-level LSTM model using six typical-day cases from the 2025 dry season. Case selection uses power-output profiles to identify days with modest daily peaking and small intraday revisions to the original power schedules. The coupled model reduces both mean absolute error (MAE) and root mean square error (RMSE) in four cases. Across all six cases, average MAE and RMSE decrease from 0.13 and 0.15 m to 0.07 and 0.08 m, respectively, and average end-of-day absolute error decreases from 0.27 to 0.03 m. The method can support the preparation of daily generation schedules for run-of-river hydropower stations under modest daily peaking and relatively stable operating conditions.