Time series compression reduces data volume while preserving essential patterns for neural network training. Critical for processing high-frequency B-Train data (1 kHz) before feeding to models like Temporal Fusion Transformer.

Problem Statement

The sampling rate of raw B-Train data is 1 kHz, meaning one second of data is 1000 points. For 10s predictions, this becomes 10,000 points, which is significant for large neural networks, especially LSTMs where computation time scales directly with number of samples.

Compression Methods

Adaptive Algorithms

Fixed-Rate Methods

  • Regular interval downsampling - introduces artifacts in high rate-of-change regions

Consequences

Non-uniform Time Axis

Adaptive downsampling creates irregularly spaced samples. Must add temporal information as features:

  • Relative time indices ()
  • Absolute time (), normalized

Applications