MISO researcher says AI ensemble LAFNET cut day‑ahead forecast errors across 36 LBAs

Dr. Gizmati of MISO presented LAFNET, an AI‑driven forecast‑fusion framework that he said ‘‘consistently achieved a lower forecast error than [the] best individual winner’’ in day‑ahead (24‑hour, hourly) load forecasting.

He told attendees the project responds to a practical problem: no single weather forecast source performs best across all regions and conditions, so the team developed a model that learns historical vendor error patterns and dynamically assigns weights to multiple weather‑driven forecasts. Dr. Gizmati said the model name LAFNET stands for Least‑Square Attention Fusion Network and that it consists of an LSTM component to capture temporal error patterns, an attention mechanism to find relevant historical situations, a fusion layer to weight vendor forecasts, and a residual correction layer to remove systematic bias.

Why it matters: accurate day‑ahead load forecasts affect unit commitment, market operations and system reliability, and reducing systematic underprediction can cut reserve shortfalls and market inefficiencies, Dr. Gizmati said.

How the team evaluated LAFNET: Dr. Gizmati described a nine‑month evaluation window (from June through the end of February in the test period) across 36 MISO local balancing authorities (LBAs). The team reported mean absolute error and mean bias error as its primary metrics. He said LAFNET outperformed the best individual vendor in many LBAs and produced substantial system‑level gains; in his summary he cited roughly 37% and 35% improvements for two of the largest LBAs and about a 51% improvement aggregated across the 36 LBAs during the test window.

Peak hours and bias reduction: Dr. Gizmati highlighted that LAFNET reduced peak‑hour error by about 20% compared with the best vendor in his examples, and that it substantially reduced systematic underprediction bias. He reported an example change in the mean bias metric at peak from figures he gave as about 1,826 megawatts down to about 1,410 megawatts, which he framed as a meaningful removal of underprediction.

Case studies and robustness: The team ran case studies on heat days and reported large day‑specific improvements: Dr. Gizmati cited an 81% improvement on Aug. 19 in the test set and a 47% improvement on Sept. 5 for the examples shown. He acknowledged the nine‑month window did not fully include a full summer and said the team will continue monitoring performance across additional seasons to confirm robustness.

Implementation details: LAFNET is implemented by MISO staff in the cloud environment (Azure), Dr. Gizmati said. He stated that MISO’s current production forecast is built on a commercial vendor software package maintained by that vendor, and that LAFNET is an in‑house layer built on top of production forecasts and vendor weather inputs.

Questions from the room: Attendees asked about forecast horizon and inputs, model architecture, and regional differences in performance. Dr. Gizmati confirmed the work focuses on day‑ahead (next 24 hours, hourly) forecasts and listed weather inputs used as temperature, wind speed, dew point/humidity and cloud cover among the main features. When Derek Wax (OCNC) asked about alternative architectures, Dr. Gizmati said LSTM is used to memorize temporal patterns but it is combined with attention, fusion and residual correction rather than used in isolation. An online questioner asked about tools; Dr. Gizmati reiterated the production model uses commercial software while the research LAFNET stack runs on Azure under MISO control. Student Freda Monchmidt asked why gains were larger in the central region; Dr. Gizmati said he suspected differences in weather‑forecast accuracy across regions but had no firm explanation.

What’s next: Dr. Gizmati said the team will continue assessment over longer time horizons and across additional seasons, with operations and risk assessment staff participating in validation and use‑case testing.