Hybrid model benchmark · live demo

預測 · Two-week rolling forecast (test period)

actualCNN-BiLSTM-Attention (JS)persistence

結果 · Metrics on these 336 hours (computed in your browser)

Try a single forecast

Forecast origin (hour of the test period): Add a gust to the last 3 input hours: m/s Shift the whole 24-h input: m/s

Next-hour forecast

–

Bars: the 24 input hours. Line: attention weight α_t over the 22 convolution steps.

Synthetic data disclaimer. This page runs a small demo network (16 filters, 16 LSTM units per direction) trained for a few epochs on synthetic SCADA-style data. It is a small version of the model from my paper, not the full model, and the data is not the paper's data, and its numbers say nothing about the published results. Persistence (next hour = this hour) is shown so you can judge the model honestly.