Overview
Real-world time series data rarely consists of a single, uniform pattern; it is usually a mixture of different underlying mechanisms, such as normal traffic versus rush hour. The Probability Pattern-Guided Time Series Forecasting (PPGF) framework addresses this complexity by reformulating forecasting as a two-part problem: first, classifying which "pattern interval" a future value belongs to, and second, predicting its exact relative offset within that interval. This approach enforces consistency between classification and regression, leading to significant accuracy gains, particularly on imbalanced or extreme data.
The Challenge of Mixed Patterns
Most standard time series forecasting (TSF) models treat historical data as a single continuous pattern. However, practical data is often messy and contains multiple internal mechanisms. For instance, solar wind speed can be slow and steady (caused by coronal holes) or fast and erratic (caused by coronal mass ejections). Because these patterns occur at different frequencies, the data is often highly imbalanced and long-tailed.
When standard models attempt to fit this mixed data, they tend to over-optimize for the majority pattern, leading to poor performance on minority or extreme patterns. While some recent methods combine a classifier (to predict the event type) with a regressor (to predict the exact value), they often lack consistency—a model might classify an event as "extreme high" but output a numerical value that falls in the "normal" range. PPGF addresses both the data imbalance and this lack of consistency.
The PPGF Framework
PPGF is an end-to-end framework that tightly couples pattern classification with numerical forecasting through several key mechanisms:
- Balanced Pattern Grouping: Instead of dividing target data into equal-width bins (which exacerbates imbalance), PPGF uses an equal-frequency grouping strategy. It divides continuous target values into $K$ non-overlapping intervals such that each interval contains roughly the same number of samples.
- Temporal Information Extractor: A shared backbone combines 1D-Convolutions for local features, a Transformer encoder for long-term dependencies, and a Gated Residual Network (GRN) for adaptive complexity to extract general temporal features.
- Probability Pattern Classifier with TCP: To improve reliability, PPGF uses True Class Probability (TCP) calibration. A dedicated confidence network (ConfidNet) estimates the reliability of extracted features. If confidence is low, the model scales the features to focus more on these "hard" samples, improving classification accuracy.
- Relative Prediction Strategy: Instead of predicting the absolute future value $Y$, the regression head predicts $\Delta Y$—the offset from the lower bound of the predicted class interval.
- Guaranteed Consistency: The final predicted value is strictly calculated by adding the predicted relative offset to the lower bound of the class interval chosen by the classifier.
Experimental Results
The authors evaluated PPGF on six diverse real-world datasets, including Power Consumption, Steel Industry Energy, Metro Traffic, Air Quality, Electricity Load, and Solar Wind Speed.
Key Findings
- State-of-the-Art Performance: PPGF outperformed 14 baseline models (including LSTM, TCN, PatchTST, and DXtreMM) in 19 out of 24 experimental configurations.
- Massive Error Reductions: On datasets with weak periodic patterns, PPGF achieved impressive RMSE reductions compared to baselines: 39% on Air Quality, 61% on Electricity Load, and 30% on Solar Wind.
- Learning Efficiency: By forcing the regression head to only predict a small offset within a known boundary, the forecasting task becomes significantly easier and more accurate than predicting global variance.
Limitations and Future Work
While PPGF shows great promise, the number of pattern groups ($K$) currently remains a manual hyperparameter requiring tuning (e.g., $K=8$ for Power Consumption vs. $K=2$ for Solar Wind). Additionally, the model can be prone to errors for samples sitting exactly on the boundary between two intervals. Future research will explore data-driven methods to automatically learn the optimal number of patterns and further refine boundary handling.
Summary updated