⚡ Multi-Modal Deep Learning for Energy Demand Forecasting

Per-zone α (weight on Baseline; 1−α goes to Chronos): ME: 0.30 | NH: 0.30 | VT: 0.80 | CT: 0.00 | RI: 0.10 | SEMA: 0.00 | WCMA: 0.05 | NEMA_BOST: 0.00

Strict-discipline backtest — at each forecast time T the model sees only data available before T. History weather: 24 HRRR f00 analyses; future weather: f01..f24 from cycle T-1 (the most recent cycle issued before T).

Last 7 days of forecasts — per-zone & overall MAPE (%)

Window: 2026-07-14T00:00 UTC → 2026-07-20T00:00 UTC · refreshed 2026-07-21T06:30 UTC

Each forecast issues a 24-hour prediction at 00:00 UTC. The baseline uses real HRRR f00 analyses for the history window (24 cycles) and HRRR f01..f24 forecasts from the cycle issued at T-1 for the future window — strict deployable backtest with no future-data leak. See About for the disclosure on the training-time future_weather mismatch.

📊 Live performance — refreshed daily at 04:00 UTC

Window: 2026-07-14 → 2026-07-20 · refreshed 2026-07-21T06:30 UTC

See the Backtest tab for the full per-zone breakdown. Numbers above are computed end-to-end on real ISO-NE per-zone load + real HRRR f00 analyses + HRRR f01..f24 forecasts (no future leakage); read the prose below for why these are higher than the offline 4.21 % headline.

About this demo

This Space runs a trained CNN-Transformer baseline on fully real, live ISO New England inputs, blended with Chronos-Bolt-mini (Amazon, 21 M params, Apache-2.0, zero-shot on demand history alone) in a per-zone weighted ensemble. It is also the case study for our workshop paper on deployment drift and inference-time ensemble adaptation under behind-the-meter (BTM) solar buildout — see the paper PDF and the v1.6 release notes in the main README.

There are two tabs:

1. Real-time forecast — every click pulls the latest 24 h of demand and HRRR weather, plus a 24 h HRRR forecast cycle, and produces a 24-hour 8-zone prediction.
2. Backtest (last 7 days) — 7 daily forecasts on the most recent 7 fully-published days, refreshed every day at 04:00 UTC by a GitHub Actions cron in new-england-real-time-power-predict-data.

⚠ Why deployment MAPE differs from the offline 4.21 % headline. The trained baseline saw 2019–2022 weather + demand at training time. Deployed today it sees 2026 weather + demand, and 3+ years of New-England-grid evolution (utility-scale BTM solar, EVs, post-COVID load patterns) has shifted the per-zone signal — most visibly in the dense southern coastal zones (RI / SEMA / WCMA). The pipeline itself is verified correct: re-running it on 2022-12-30 reproduces the cluster's 6.54 % MAPE to within 0.13 percentage points (we observe 6.41 %, see Live performance validation in the README). The gap between the headline and the deployed numbers is dominated by honest training→deployment drift, not bugs.

What's real (everything)

⁽¹⁾ One small caveat: the APCP_1hr (1-hour accumulated precipitation) channel is zero-filled for HRRR forecast hours because the search regex that worked at training time matches a different accumulation window than the operational forecast files publish. Six of the seven weather channels are real; APCP_1hr's training mean is small (~0.4 mm/h) so after z-scoring this approximates the training mean, with measured impact <1 % MAPE.

The bundled 2022 sample CSVs are kept ONLY as a final fallback for when the live ISO-NE / HRRR endpoints are unreachable.

Strict-discipline backtest

For each daily forecast at time T (the last 7 days at 00:00 UTC each):

• Demand history for hours [T-24, T-1] comes from the public 5-min zonal feed
• Weather history is 24 HRRR f00 analyses, one per valid hour [T-24, T-1]
• Weather forecast is HRRR cycle (T-1)'s f01..f24 — i.e. the most recent forecast that was issued before T, with valid hours [T, T+23]
• Truth for MAPE is the ISO-NE per-zone load for [T, T+23]

In particular no future analyses are used — every forecast at T sees only data that would have been available at time T, matching what a real deployment would do.

Disclosure: training-time future_weather mismatch

The trained baseline saw HRRR f00 analyses for both history AND future windows during training (i.e. the model was given the actual weather that occurred during the prediction window as an input feature). This is a form of supervised-learning-with-privileged-information that the report acknowledges in its Part-2 future-weather ablation.

At deployment we cannot use future analyses (they don't exist for the future yet), so we substitute HRRR forecasts (f01..f24) issued at the cycle just before the forecast time. The model therefore sees a slightly out-of-distribution input for the future window. This Space measures the deployable accuracy honestly. The Chronos-Bolt-mini ensemble path partially compensates because Chronos doesn't use weather at all.

Per-zone allocation — actually per-zone now

Earlier prototypes of this demo used a fixed proportion vector to split the system total (from the EIA Open Data API) into 8 zones, which made the per-zone view cosmetic. The current Space pulls true per-zone load from ISO-NE's 5-minute estimated zonal feed, so per-zone forecasts are real.

First-call latency

The first Live tab click triggers 24 HRRR analysis cycles + 24 forecast hours from AWS S3 (parallel-fetched, cached at /tmp/hrrr_cache/) plus a one-time Chronos-Bolt-mini load (80 MB from HuggingFace Hub). Expect ~3-5 minutes on the very first click of a fresh Space instance, then ~10-30 s on subsequent clicks within the same uptime session. The Backtest tab is instant — its data ships pre-computed from the data repo.

Per-zone ensemble weights

Per-zone $\alpha_z$ (shown beneath the chart) blends baseline and Chronos:

$$\hat y_z = \alpha_z \cdot \hat y_z^{\text{baseline}} + (1 - \alpha_z) \cdot \hat y_z^{\text{Chronos}}$$

$\alpha_z$ values come from a grid search on a 14-day validation window in 2022. See the Foundation-model ensemble section of the report for the full ablation across context length, model size, and quantile aggregation.

Links

• 📄 Final report (PDF)
• 💻 Main code repo
• 🤖 Auxiliary data repo (cron-refreshed backtest data)
• 👤 Author: Pang Liu · Independent Researcher · jeff.pang.liu@gmail.com

Figures from the report