Does knowing where the sun is help forecast Delhi's demand?

Trees beat the LSTM, as in the paper. LightGBM reaches 0.87% against the LSTM's 1.06% on the same windows, from features built out of the same 24 hours of history.

For trees, the sun adds nothing the calendar does not already say. LightGBM scores 0.87% without the sun features and 0.87% with them. Hour of day and day of year fix where the sun is, and trees can learn that mapping on their own. A linear model cannot build that curve from sines and cosines, and there the sun features do help a little: 1.51% to 1.49%. But the gain is not in daylight hours (0.00 points there), so it reads as a better daily curve for the linear model, not as a solar effect.

Not the paper's number. The paper reports 0.77% MAPE on its own 18 months of data and setup. These results use different data and should not be compared with it; what carries over is the method and the ranking of the models.

Same problem as the LSTM. At each origin, load and weather are known up to five minutes before it; forecast the next 12 five-minute steps. Same five-minute grid, gap handling and breaks at impossible jumps, same 2024 test year.

Features. From the previous 24 hours only: the last load, its changes over the last 5 minutes to 4 hours, the 1-hour and 24-hour means, the load at the target time yesterday and how it moved from now to then, temperature, humidity and the temperature trend. From the calendar at the target time: hour, weekday and day of year, as sines and cosines. The sun group adds the sine of the sun's elevation at the target time and its change since the last reading.

Models. One model per horizon predicts the change from the last reading. Ridge regression (alpha 1) on standardised features. LightGBM with 63 leaves and learning rate 0.05; its number of trees per horizon is set by early stopping on 2023 while training on the years before, then the model is refit on all training years. The LSTM is the one on its own page, run through the same ONNX file the browser uses.