Modelling Seasonality and Trends in Daily Rainfall Data

Peter M. WilliamsNeurIPS 1997

This paper, published by Peter M. Williams in the mid-90s, applies neural networks to a classic meteorological challenge: making sense of the "noisy" nature of daily rainfall to find underlying climate patterns. By predicting the parameters of a Gamma distribution and integrating over possible past weather states, the model successfully extracts smooth seasonal cycles and long-term climate trends from decades of volatile data.

The Challenge of Rainfall Data

Rainfall data is notoriously difficult to model because it is "stochastic" (random) and "non-stationary" (changing over time). Traditional methods often separate the occurrence process (did it rain?) from the amount process (how much did it rain?).

The challenge lies in capturing the dependencies: how does the weather over the last week affect today’s probability of a downpour? Furthermore, how do we separate the "noise" of a single storm from the "signal" of a changing climate or a seasonal cycle? The author aims to create a unified model that can handle these short-term dependencies while revealing long-term trends.

Methodology

The author treats rainfall as a mixture of a discrete component (zero rain) and a continuous component (a Gamma distribution for rainy days).

  • Neural Network as Parameter Predictor: Instead of predicting a specific amount of rain, the neural network outputs three distributional parameters: $\alpha$ (probability of rain), $\nu$ (shape), and $\theta$ (scale). This allows the model to define the entire "shape" of uncertainty for any given day.
  • Inputs: The model uses a 10-day history of rain occurrence (binary), cyclic variables ($\sin$ and $\cos$ of the time of year) to capture seasonality, and a linear time variable to detect long-term trends.
Conditional probability and mean for Falmer
  • Bayesian Regularization: To prevent the network from "overfitting" to specific noisy years, the author uses Bayesian regularization and pruning. This automatically simplifies the model, effectively "turning off" unnecessary connections.
  • The Integration Trick: To extract a clean seasonal curve, the author "integrates out" the short-term noise. By mathematically averaging the model’s predictions over all $2^{10}$ possible 10-day weather histories, they produce a smooth marginal expectation for any day of the year.

Datasets

The study utilizes roughly 40 years of daily rainfall data (approx. 15,000 observations per station) from 10 different weather stations:

  • South East England: Cromptons, Falmer, Kemsing, Petworth, and Rotherfield.
  • Central Italy: Monte Oliveto, Pisticci, Pomarico, Siena, and Taverno d’Arbia.

Key Insights and Results

  • Adaptive Seasonality: Traditional models use a fixed number of harmonics (Fourier series) to model seasons. By providing $\sin$ and $\cos$ as inputs to a neural network, the model adaptively determines the complexity of the seasonal cycle.
  • Rain Persistence: The model confirms that the probability of rain increases if it has rained recently but tends to level off after 5–7 consecutive rainy days. This validates the choice of a 10-day "memory" for the model.
  • Joint Modeling: By predicting the parameters of a Gamma distribution, the model captures the reality that smaller amounts of rain are generally more likely than large amounts, but the "scale" of that likelihood changes based on recent history.
  • Model Efficiency: Through pruning, the network reduced its complexity from 343 potential parameters down to an average of just 65, indicating a highly efficient representation of the data.

In Pomarico, Italy, the model revealed a significant downward trend in mean rainfall from 1955 to 1985. Interestingly, the frequency of rain stayed roughly the same, but the amount of rain per event decreased—particularly during the winter months.

Probability of rain and mean/standard deviation trends for Pomarico
  • Ergodicity: The author demonstrated that the model is "ergodic," meaning the predicted probability of rain becomes independent of the starting conditions after about 50 days.

Limitations and Future Work

  • Simplified History: To keep the mathematics of integration tractable, the model only considers whether it rained in the past 10 days, not the exact amount that fell. The author acknowledges this is a simplification.
  • Single Gamma Distribution: The continuous component uses a single Gamma distribution. The author suggests that a "mixture of Gammas" could provide even more flexibility for complex weather patterns.
  • Spatial Independence: The paper models each station in isolation. Future work could potentially incorporate spatial correlations between neighboring stations to improve predictive power.

Summary updated

BibTeX
@inproceedings{williams1997modelling,
  author = {Peter M. Williams},
  title = {{Modelling Seasonality and Trends in Daily Rainfall Data}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1997},
}
Figure