Exogenous-Aware Spatio-Temporal Demand Forecasting

This direction is merged into another direction: exogenous variable integration. Its last version stays readable and citable here.

Version 1 of 130 papers30 conceptsfrozen 2026-10-04

Predicting the demand for resources—whether electricity, natural gas, or ride-hailing services—is a fundamental challenge in urban planning and grid management. While traditional models rely heavily on historical usage patterns (autoregression), modern research has shifted toward exogenous-aware forecasting. This approach recognizes that demand is not just a function of the past, but a response to a complex web of external drivers like weather, economic shifts, and social behavior.

The Failure of Historical Autoregression

The "historical load" of a power grid or a traffic network is inherently volatile. As Li Ran and colleagues argue in their Short-Term Bus Load Matching and Forecasting Model, models that look only at past data fail to capture the "why" behind sudden spikes or dips. If a heatwave begins, historical data from a cool week is a poor predictor of tomorrow’s air conditioning load.

This limitation is echoed by Grzegorz Dudek and Pawel Pelka in their work on N-BEATS for electricity forecasting, where they note that autonomous models are blind to external influences. To move beyond simple curve-fitting, researchers are integrating exogenous variables: non-load data streams that correlate with demand.

Integrating Multi-Modal Exogenous Signals

The core intuition of this research direction is that demand is a "proxy" for human and environmental activity. To model it accurately, we must fuse diverse data types:

  • Meteorological Data: Temperature, humidity, and wind speed are primary drivers of energy use. Patrick Emami and the NREL team demonstrated this with BuildingsBench, showing that weather covariates are essential for generalizing models across different climates.
  • Socio-Economic Indicators: Public holidays, fuel prices, and local events shift demand in ways that physical laws cannot explain. Sui Zhang and Baoyue Wang show that combining these with meteorological data allows for much more precise P2P energy trading.
  • Spatial Context: Demand at one location is often influenced by its neighbors. Zongbao Zhang uses Multiscale Spatio-Temporal Enhanced Models (MSTEM) to capture how EV charging demand ripples across a city's geography.

Architectural Fusion of Time and Space

How do we actually combine these signals? The research points to a "divide and conquer" architectural strategy.

First, models must handle temporal dependencies. Wenshuai Jiang and others utilize multilayer dilated LSTMs with attention to focus on critical historical moments while maintaining a long-range view. Second, they must handle spatial dependencies. Shilu Yuan and Xiaoyu Li introduced STPro, which identifies "spatio-temporal prototypes"—shared patterns across different parts of a city—to predict Origin-Destination demand.

The most advanced frameworks, like GESME-Net by Md. Hishamur Rahman, use a Gated Ensemble of Experts. This allows the model to dynamically weigh different inputs (e.g., giving more weight to weather during a storm and more to historical trends during a calm week).

Quantifying Uncertainty and Bias

Because exogenous variables like weather forecasts are themselves imperfect, a single "point prediction" is often insufficient. Dahua Gan and Chongqing Kang address this through embedding-based quantile regression, which predicts a range of possible outcomes rather than a single number.

In complex terrains like the Arctic, Finn Dag Hovem found that deep learning models can actually auto-correct biases in regional weather models by learning the specific physical constraints of a local site. This suggests that the model isn't just using exogenous data—it's learning the relationship between the environment and the resource.

Interpreting the Black Box

As these models become more complex, understanding why a prediction was made becomes vital for decision-makers. Yifan Jiang and Xuan Song developed TrafPS, a system that uses Shapley values to decompose deep-learning forecasts. This allows users to see exactly how much a specific exogenous factor—like a nearby accident or a sudden rainstorm—contributed to a predicted traffic jam.

By moving from simple autoregression to exogenous-aware, spatio-temporal architectures, the field is evolving toward "foundational" utility models. As seen in the work of Xinxing Zhou on natural gas demand, these models are beginning to generalize across different industries and regions, treating demand as a universal language of human-environment interaction.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_exogenous_aware_spatio_temporal_demand_f_4280f804,
  title = {Exogenous-Aware Spatio-Temporal Demand Forecasting},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/exogenous-aware-spatio-temporal-demand-forecasting/rdr_4280f804a97e4103b08abd203d75df4a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_exogenous_aware_spatio_temporal_demand_f_4280f804_v1,
  title = {Exogenous-Aware Spatio-Temporal Demand Forecasting},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/exogenous-aware-spatio-temporal-demand-forecasting/rdr_4280f804a97e4103b08abd203d75df4a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}