Wind Pressure Cp Forecasting — BWU Research

Deep-learning and classical forecasting of wind pressure coefficients ($C_p$) on tall buildings using the TPU BDH benchmark.

→ Project Website · Interactive Dashboard

---

Headline Results

Long-horizon (h=500) numbers below are per-step error aggregated over the full 146,880-window test set — the same regime for every model. (An earlier version scored the autoregressive models on a single 500-step trajectory, which inflated the naive baseline to RMSE 0.118 / R² 0.97; that was a measurement artifact — see evaluate_long_horizon.py.)

Model	Paradigm	RMSE h=500	R² h=500	vs Naive
PatchTST	Direct multi-step	0.175	0.85	beats
LSTM-direct	Direct multi-step	0.179	0.84	beats
Random Forest	Autoregressive	0.212	0.78	beats (marginal)
XGBoost	Autoregressive	0.217	0.77	tie
Naive persistence	—	0.218	0.77	baseline
LSTM autoregressive	Autoregressive	0.270	0.65	slightly worse
GRU autoregressive	Autoregressive	0.273	0.64	slightly worse
Ridge	Autoregressive	0.281	0.62	slightly worse (stable)
TCN autoregressive	Autoregressive	0.411	0.19	collapses

(Classical rows re-fit at window-stride 40; naive = 680-trajectory value. The exact 146,880-window naive is 0.231, which the two direct models also beat. The exact tree-vs-naive margin is sensitive to training-set size.)

Key findings:

1. The direct multi-step models clearly beat naive persistence at h=500 (PatchTST 0.175, LSTM-direct 0.179 vs 0.218–0.231) — by changing only the training objective, not the architecture.
2. The tree-based autoregressive models sit right at the naive level — Random Forest (0.212) marginally beats it, XGBoost (0.217) ties. The exact margin moves with training-set size, which is itself the finding: at long horizon these models hit the predictability ceiling around persistence, they do not clear it.
3. The earlier "catastrophic collapse" of autoregressive deep models (reported R² = −6.18) and "Ridge diverges to ~10⁴" were both single-trajectory artifacts. Over 680 trajectories LSTM/GRU reach R² ≈ 0.65, Ridge is stable (0.281), and only TCN is genuinely unstable.
4. Despite winning on RMSE/R², the predicted PSD of the direct models stays 60–75× below the true-signal PSD at all frequencies — MSE training collapses to the conditional mean, so the forecasts remain unusable for spectrum-dependent wind-engineering quantities (fatigue, peak factor).

Repo Structure

Agent_Test/      ← all models, training scripts, results
Agent_Papers/    ← WPTSE-Net synthesis model
AI_Agent/        ← Streamlit dashboard + multi-agent system
Data/            ← TPU BDH raw data (gitignored)
docs/            ← GitHub Pages website

📚 All docs in one place: see DOCS.md (documentation index).

Reproduce

conda activate ML_Cesar
$env:KMP_DUPLICATE_LIB_OK = "TRUE"
cd Agent_Test
python train_all.py --scope all          # Round 3 (full)
python analyze_horizons.py --scope all   # multi-horizon metrics + plots

Dashboard (local)

pip install -r requirements.txt
streamlit run AI_Agent/streamlit_app.py