AI-based smart energy consumption prediction for sustainable buildings using IoT data

Two complementary evidence layers are investigated for this study to look at next-hour building electricity forecasting. The first is a commercial building case study showing a 14 month monitoring experience of a 12,400 square metre office building in Bengaluru, India. The statistics revealed in the past include an RMSE of 3.27 kWh, a MAE of 2.41 kWh, a MAPE of 4.83%, a CoD of 0.971, and a median edge inference time of 38 ms. Unfortunately, the original telemetry, feature dictionary, checkpoints and residuals at the time stamp level were not available, and therefore these values are not reproduced independently but rather site records from the past. The second evidence layer is a fully executable external validation based on 19,735 ten-minute observations from public Appliances Energy Prediction dataset. Following these steps, 469 held-out hours were used for testing 11 forecasting approaches. Eleven forecasting approaches were compared on 469 held-out hours after a series of steps of hourly aggregation, causal lag construction, training-only mutual-information selection and chronological partitioning. Random Forest produced the lowest root mean square error of 324.71 W-hours, Ridge regression produced the lowest mean absolute error of 183.58 W-hours, and the convolutional bidirectional recurrent model produced 330.96 W-hours root mean square error. The attention augmented hybrid, on the other hand, proved unstable for three random seeds (with a mean RMSE of \:425.51±\:59.91 Wh). A daily dependence adjusted Diebold-Mariano test failed to find superiority of the three-seed attention forecast. The results demonstrate that the sensor-to-edge-to-control paradigm is promising at the original application site but that the attention mechanism is not inherently the best and should not be the focus of the innovation.