Forecasting¶
keras_climate.forecasting — multivariate time-series forecasting.
None of the models in this family have a publicly available general-purpose pretrained checkpoint (the official repos generally only ship per-benchmark-dataset training scripts/configs, not reusable backbones) — each is validated against a from-scratch synthetic PyTorch reference of its own architecture instead. See Weight Porting for how those round-trip tests work.
PatchTST¶
Channel-independent patch-based transformer (Nie et al. 2023) for long-horizon multivariate forecasting. Each variate (e.g. a climate station's temperature, humidity, pressure channel) is patched and encoded independently by a shared transformer, then linearly projected to the forecast horizon.
padding_patch="end" (the official default) replicates the last timestep
stride times before patching, so the final stride-sized tail of the
series still contributes one more patch instead of being dropped. Pass
padding_patch=None for the simpler no-padding variant.
TimesNet¶
Discovers dominant periodicities in a series via FFT (Wu et al. 2023), reshapes the 1D series into 2D tensors (period × num_periods) for each dominant period, applies Inception-style 2D convolutions to capture both intra- and inter-period variation, then adaptively fuses the multi-period representations by their softmax-normalized FFT amplitude weight. Well suited to seasonal climate variables (diurnal / annual cycles).
TFT (Temporal Fusion Transformer)¶
LSTM encoder-decoder for local processing (Lim et al. 2021), Gated Residual Networks for variable selection and static-covariate enrichment, and interpretable multi-head attention (value projection shared across heads) over the full lookback+horizon for long-range dependencies. A strong choice for multi-horizon forecasting with known future inputs (e.g. forecast weather drivers) and static metadata (e.g. station location/elevation).
Inputs: past_inputs (historical observed drivers), future_inputs
(known future drivers, e.g. an NWP forecast), static_inputs (station
id/location/etc). Output: quantile forecasts over the decoder horizon.
This module deliberately omits the paper's static_context_selection
vector (which would condition the variable-selection GRNs with an extra
context input) — see Implementation Notes for why.
DLinear¶
Decomposes the input into a trend (moving-average) and a seasonal (residual) component (Zeng et al. 2022), each passed through its own single Linear layer over the time axis, then summed — a deliberately minimal baseline that outperformed several contemporary Transformer forecasters on common long-horizon benchmarks.
N-BEATS¶
A deep stack of fully-connected blocks (Oreshkin et al. 2020). Each block maps the lookback window to a small set of "theta" coefficients that parameterize an interpretable basis expansion (generic, polynomial trend, or Fourier seasonality), producing both a "backcast" (subtracted from the residual seen by later blocks) and a "forecast" (added into the running total) — the paper's "doubly residual stacking". Operates on a single (univariate) series per call; apply per-channel for multivariate series.
Informer¶
An efficient long-sequence Transformer forecaster (Zhou et al. 2021, AAAI Best Paper) built around: "self-attention distilling" (halving the sequence length between encoder layers via a strided conv, so a deep encoder stays tractable over long lookback windows), and a generative-style decoder that predicts the whole forecast horizon in one forward pass rather than autoregressively. This module uses standard full scaled-dot-product attention in place of the paper's ProbSparse query-sampling — a computational optimization (an approximation of full attention) rather than a different learned function, so this only changes speed characteristics.
Autoformer¶
A decomposition-based Transformer forecaster (Wu et al. 2021, NeurIPS). Series-decomposition blocks (moving-average trend/seasonal split) are embedded throughout the encoder and decoder rather than applied once up front. Self-attention is replaced by an Auto-Correlation mechanism that discovers period-based dependencies via FFT. This module computes a mathematically-equivalent-in-spirit full (non-top-k) softmax-weighted circular convolution in place of the official mechanism's dynamic top-k lag selection + circular shift, which don't translate cleanly to a static, backend-agnostic Keras graph — see Implementation Notes.