Does Urban Energy Demand Really Flow? A Falsifiable Test for Continuous Demand-Field Models

A small line of work on urban electricity models a city's demand as a fluid. Smart-meter readings are smoothed into a continuous map of demand. Two maps from different times of day are subtracted. The spatial slope of that difference is then drawn as arrows showing load "moving" from business districts to residential suburbs each evening. The pictures are striking, but nobody has checked what the arrows mean. No sensor measures kilowatt-hours travelling down a street. The published models are scored only by squared error on the reconstructed loads, and every one of them is trained and tested on a single district of Shanghai. This proposal asks one question. Does the inferred flow field carry predictive information that a location's own history does not? It answers with a metric that needs no ground-truth velocities, calibrated on synthetic cities where the true answer is known, and applied across time periods, spatial regions and datasets.

The core idea fits in two sentences. If demand really migrates, then the recent history of places "upstream" of a location, as defined by the inferred arrows, should improve forecasts at that location beyond its own seasonal history and beyond equally distant places that are not upstream. If it does not, the flow field is a relabelled difference map, and the fluid picture is decoration.

Why this question, and why now

The fluid pipeline has three stages. First, kernel density estimation (a smoothing step that spreads each meter's reading over a small neighbourhood) builds demand maps from Shanghai Pudong meters. Next come differential maps between time windows. Last, a potential-flow vector field is computed as the negative gradient of the difference. This pipeline is used in A flexible potential-flow model based high resolution spatiotemporal energy demand forecasting framework, Understanding energy demand behaviors through spatio-temporal smart meter data analysis and the VAP tool VAP: A Visual Analysis Tool for Energy Consumption Spatio-temporal Pattern Discovery. Interpretable spatiotemporal urban energy forecasting forecasts the density maps as image sequences with recurrent convolutional networks. It reports a 7-day mean squared error (MSE) of 6.032 and concedes that the model may need retraining for cities with different climate, morphology or building stock.

These papers motivate the proposal. Two observations constrain it.

The first is a simple mathematical point. A gradient field computed from a difference map contains no information beyond that map. The arrows are fully determined by where demand rose and fell, so they cannot by themselves distinguish "commuters carried load from A to B" from "A's businesses closed while B's homes switched on independently."

The second is empirical. A Critical Audit of Spatiotemporal Forecasting Benchmark Datasets and Baselines audited traffic, delivery, disease and page-view benchmarks. It showed that their predictive signal is almost entirely temporal. Neighbours add little once a node's own autoregression is accounted for, and a tuned seasonal autoregressive model (SARIMA) beats most graph networks on WikiMaths. The potential-flow paper itself reports that near-zero "idle" and "energy-saving" customers suffer high relative error under density projection, and it suggests ARIMA may do better there. Neither energy paper compared against a tuned SARIMA baseline.

Together these observations sharpen the opportunity. The field needs a quantitative, falsifiable score for the flow representation itself, and a check of whether that score survives outside the data it was fitted on.

Core Claim

We claim that the physical content of an inferred demand-flow field can be measured by its directional predictive gain. This is how much the flow-designated upstream neighbourhood of a location improves held-out forecasts there, relative to a distance-matched set of non-upstream neighbours, after a tuned seasonal model of the location's own history has been removed. Writing $e_{\text{own}}$ for held-out error using only the location's own history, $e_{\text{up}}$ for error when upstream-neighbour history is added, and $e_{\text{ctrl}}$ for error when an equally sized, equally distant control neighbourhood is added instead, we define

$$ G = \frac{e_{\text{ctrl}} - e_{\text{up}}}{e_{\text{own}}}, $$

so $G > 0$ means the arrows point to places whose past genuinely explains the future, beyond what generic spatial smoothing provides. The control term matters. Kernel smoothing and plain spatial correlation will make any nearby history somewhat useful, and only the excess attributable to direction tests the migration claim.

Two design rules make $G$ honest. First, it is always computed on individual meters that were held out of the density estimate, so the smoothing kernel cannot leak a target's own reading into its "neighbours." Second, it is reported per consumption class and per region, not only as a city-wide average.

Research Program

The program has three parts.

The first part is calibration on synthetic cities, following the logic of Comparing Clustering Approaches for Smart Meter Time Series: Investigating the Influence of Dataset Properties on Performance. That paper built expert-designed synthetic load shapes to escape circular validity indices, and it showed that synthetic rankings tracked expert-labelled real data (r = 0.939). We borrow the method, not the load shapes. We generate cities in which a commuter population moves along a known origin-destination pattern and carries its demand with it, and we generate matched "null" cities with identical marginal diurnal curves at every location but no movement.

The prediction is that potential-flow arrow maps look similar in both kinds of city, while $G$ separates them. We will also sweep noise, meter density and kernel bandwidth, to find the regimes in which $G$ has statistical power. If $G$ cannot tell the two cities apart, the metric is abandoned before real data are touched.

The second part applies the metric to Shanghai Pudong, the only geolocated energy cohort in our material, at the resolutions the three Shanghai papers used. We test three kinds of shift: a temporal shift (fit on 2017, test on 2018, mirroring the potential-flow paper's split), a spatial-block shift (hold out contiguous sub-regions of the district), and a seasonal shift (shoulder months against peak months, where the potential-flow paper found an unexpected "deviation pattern").

The third part is cross-dataset validation. Here we must be candid: the research file lists no second smart-meter dataset with meter coordinates. Low Carbon London and the IEEE energy-prediction set carry sociodemographic or dwelling metadata but, as described, no geolocation. We therefore do cross-dataset work in two ways.

First, we apply the identical protocol to geolocated demand-like fields where movement is physically plausible and the audit already gives a prior: the four LargeST California regions, METR-LA against PEMS-BAY, and the New York, Chicago and Washington bike-sharing grids used by Cross-City Transfer Learning for Deep Spatio-Temporal Prediction. If $G$ is positive for bike pickups, which physically do move, but near zero for traffic speed, that is evidence that the metric reads real transport rather than smoothing artefacts.

Second, we will seek a geolocated release of a second utility cohort. This is a precondition risk we state openly, not a result we assume.

Separately, we test whether fluid-model forecasters trained on Shanghai hold their error when evaluated on held-out blocks and years. The non-spatial energy sets, such as the IEEE set and Low Carbon London, serve only as a check that our SARIMA and gradient-boosting baselines are tuned to published strength.

Baselines

The naive reference is seasonal persistence: tomorrow's value at a meter equals the value one week earlier. Every method must beat it on the same held-out meters and horizons, or no spatial claim is considered.

The first strong competitor is a per-meter SARIMA with orders chosen by information criterion on the training period. The audit shows this family is often the real state of the art on spatio-temporal data, and no energy-flow paper has faced it. The audit's own hybrid, a recurrent graph network trained on SARIMA residuals, reports MAE 3.32 on METR-LA A Critical Audit of Spatiotemporal Forecasting Benchmark Datasets and Baselines. We rerun it on each of our datasets with the same tuning budget as the fluid models. It is the strongest version of "space helps a little, on top of time."

The second competitor is a spatially unaware gradient-boosted model (LightGBM) with calendar, weather and lag features. This is the approach behind the IEEE energy-prediction winner, which reports total relative absolute error 0.6471 on that competition's data The Energy Prediction Smart-Meter Dataset: Analysis of Previous Competitions and Beyond. That figure is not comparable to Shanghai numbers, so we refit LightGBM ourselves.

The fluid models themselves are the SA-ConvLSTM and PredRNN density forecasters of Interpretable spatiotemporal urban energy forecasting, with 7-day MSE 6.032 reported, and the clustering-based LSTM of the potential-flow paper, with RMSE 9446.15 reported. We reimplement both, because the published numbers are on different resolutions and splits. Beating them means lower held-out-meter error under the spatial-block and year-shift splits, not only on their original splits.

For transfer, we include STPDE from Decoupling Universal Laws and Environmental Heterogeneity: A Physics-Inspired Framework for Robust Spatio-Temporal Forecasting. STPDE separates a shared diffusion law from learned local environment factors and reports a 7.2% few-shot cross-city improvement on LargeST. We add RegionTrans, which reports up to 10.7% RMSE reduction over ST-ResNet on the bike grids. All models get equal hyperparameter-search budgets, equal wall-clock limits per dataset, and identical held-out meters.

The ablation that isolates the key ingredient is the control inside $G$ itself. We replace flow-designated upstream neighbours with neighbours at matched distance chosen from directly downstream, orthogonal, or random bearings, and we recompute. A second ablation reverses every arrow. If reversed arrows score as well as true ones, the direction carries no information. Beating the ablation means $G$ significantly above zero, with a block bootstrap over days and spatial blocks, for true arrows and not for reversed or random ones.

Why not just…?

Why not just fit SARIMA and stop? This is the real threat, and the proposal is built to settle it rather than argue with it. The audit's evidence that spatial signal is marginal on five benchmarks, together with the potential-flow paper's own admission about sparse loads, makes it plausible that $G \approx 0$ across most of Shanghai. That outcome is a publishable, useful negative result. It is reported per consumption class, because the likely finding is that directional gain, if it exists at all, concentrates in commercial-to-residential transition zones during evening hours.

Why not just report MSE of each forecaster on another city? Field-level error rewards smooth predictions. It cannot tell whether the arrows mean anything, since a model with no notion of migration can reproduce the difference maps and therefore the arrows. Cross-city error will still be reported. But without the directional control, a good transfer score would not validate the fluid interpretation that these papers use to justify interpretability and substation-risk analyses.

Why not just adopt STPDE for energy? STPDE models only diffusion, the spreading of a quantity from high to low. Its cross-city gains show robustness, not that transport occurs. We run it as a baseline and also score its implied field with $G$. If STPDE transfers well while $G$ stays near zero, that says generalization comes from temporal and local structure, not from flow.

Why not estimate flow with optimal transport between the two maps instead of a gradient? Optimal transport (a method that finds the cheapest way to move one distribution of mass into another) always returns some transport plan, so on its own it is equally unfalsifiable. We include it as an alternative flow estimator scored by the same $G$. If it beats gradient arrows on the synthetic migration cities, that is a concrete recommendation to the field.

Decisive Checks

The work succeeds or fails on four checks. First, on synthetic cities $G$ must separate migration cities from null cities while arrow maps do not; if it cannot, the metric is discarded. Second, on Shanghai we test whether any fluid forecaster beats per-meter SARIMA on held-out meters under the year-shift and spatial-block splits with a matched tuning budget. Third, true arrows must score higher $G$ than reversed and random bearings, with the gain located where the published maps claim evening migration. Fourth, $G$ should come out positive on bike-sharing grids, where people physically move, and near zero on highway speeds, which is consistent with the audit. A positive result on all four would give the fluid view of demand its first quantitative footing. A null result on the second and third would tell grid planners that the arrows are a visualization of change, not evidence of movement.