Sounds like a classic case of turning noise into signal. I’d start by treating the fire data as a high‑frequency time series, then apply a combination of moving‑average smoothing and an exponential‑weighted model to capture the rapid growth phases. Once the baseline is clear, use robust anomaly detection—like a median‑based threshold or a Gaussian mixture—to flag the sudden spikes that correspond to lightning strikes or sudden wind shifts. The trick is to keep the models interpretable; a simple ARIMA or even a piecewise linear fit often reveals the underlying trend better than a black‑box neural net, especially when you’re trying to explain the why behind those “flame” outliers. Once you’ve identified the spike drivers, you can build a predictive framework that learns from the same patterns in your other datasets, turning those unpredictable bursts into actionable insights.

You’re right—wildfires are the ultimate outlier generator. The key is to keep the model lightweight enough for real‑time updates, then add a safety net: a simple threshold on residuals that triggers a manual check or an automated suppression rule. That way the math can flag the spike and you can act before the fire turns the data into ash.

Yeah, keep the guardrail tight—simple rules that catch a sudden residual spike will let you intervene before the pattern explodes. Trust the numbers, but don’t let the fire write the whole story.