🎓 Lesson 7
D4
Monte Carlo Dropout for Wind Forecast Uncertainty
Monte Carlo Dropout is a simple trick that turns a regular neural network into a tool that gives not just one wind speed prediction, but many possible predictions—showing how uncertain the forecast really is.
🎯 Learning Objectives
- ✓ Explain how Monte Carlo Dropout transforms deterministic deep learning models into probabilistic forecasters
- ✓ Apply Monte Carlo Dropout to a trained LSTM wind forecasting model to generate prediction intervals
- ✓ Analyze the calibration of uncertainty estimates using reliability diagrams and Brier scores
- ✓ Design a validation protocol to assess whether predicted uncertainties align with observed forecast errors
📖 Why This Matters
In open-pit mining operations, accurate short-term wind forecasts are critical for safe blast timing—excessive wind disperses explosive fumes and compromises flyrock control. Traditional point forecasts hide uncertainty; a 12 m/s prediction could be ±3 m/s or ±0.5 m/s—but you won’t know unless the model tells you. Monte Carlo Dropout delivers quantified uncertainty *for free*, using only inference-time computation—no new hardware, no retraining. That makes it indispensable for AI-augmented blast scheduling systems operating under ISO 50001 and ISEE blast safety guidelines.
📘 Core Principles
At its heart, Monte Carlo Dropout leverages the fact that applying dropout (random neuron deactivation) at test time simulates sampling from an approximate Bayesian posterior over network weights. Each forward pass yields a different output—like a virtual ensemble—whose empirical distribution captures both aleatoric (data) and epistemic (model) uncertainty. For wind forecasting, this manifests as prediction intervals (e.g., 5th–95th percentile) around each hourly wind speed estimate. Crucially, unlike true Bayesian methods, it requires no variational inference or MCMC—just repeated inference with dropout enabled (p = 0.1–0.3), making it deployable on edge devices near blast sites.
📐 Predictive Mean and Standard Deviation
The Monte Carlo estimate of predictive uncertainty uses T stochastic forward passes. The mean prediction represents the expected wind speed; the standard deviation quantifies model confidence. Calibration ensures that, e.g., 90% of true observations fall within the 5th–95th percentile interval.
Monte Carlo Predictive Statistics
μ̂ = (1/T) Σ_{t=1}^T f_θ(x; ω_t), σ̂² = (1/(T−1)) Σ_{t=1}^T (f_θ(x; ω_t) − μ̂)²Estimates the predictive mean and variance of wind speed by averaging T stochastic forward passes through a dropout-enabled network.
Variables:
| Symbol | Name | Unit | Description |
|---|---|---|---|
| μ̂ | Predictive mean wind speed | m/s | Expected value of the wind speed forecast across Monte Carlo samples |
| σ̂ | Predictive standard deviation | m/s | Empirical measure of forecast uncertainty (epistemic + aleatoric) |
| T | Number of Monte Carlo samples | dimensionless | Typically 20–100; higher T improves uncertainty resolution but increases latency |
| f_θ(x; ω_t) | Network output for sample t | m/s | Wind speed prediction under random dropout mask ω_t applied to weights θ |
Typical Ranges:
Mine-site short-term forecasting (1–6 h): 0.3 – 2.1 m/s
Offshore wind farm forecasting: 0.8 – 3.5 m/s
💡 Worked Example
Problem: A trained LSTM wind model (dropout rate p = 0.2) performs T = 50 forward passes at t+6h. Outputs: [11.2, 10.8, 11.5, ..., 12.1] m/s (50 values).
1.
Step 1: Compute sample mean: μ̂ = (1/50) Σ yₜ
2.
Step 2: Compute sample standard deviation: σ̂ = √[(1/(50−1)) Σ (yₜ − μ̂)²]
3.
Step 3: Report 90% prediction interval: [μ̂ − 1.645·σ̂, μ̂ + 1.645·σ̂]
Answer:
The result is μ̂ = 11.42 m/s, σ̂ = 0.43 m/s → 90% PI = [10.71, 12.13] m/s, which falls within the safe operational range for blast initiation (<15 m/s with ±1.5 m/s tolerance).
🏗️ Real-World Application
At Newmont’s Boddington Mine (Western Australia), an AI-augmented blast scheduling system integrates a 3-layer LSTM with Monte Carlo Dropout (p = 0.15) trained on 2-year mesoscale WRF outputs and on-site cup anemometer data (1 Hz, 10 m height). During a July 2023 blast window, the system issued a 6-h-ahead forecast of 9.3 ± 0.8