When a utility pays households to cut electricity use during a peak hour, it has to answer two questions afterward: how much load did the payment actually remove, and how much of that load came back as a "rebound" once the event ended? Today these questions are answered in one of two ways. Operators use simple regressions, which are auditable but miss the physics: buildings and batteries store energy and release it later. Or they use black-box forecasters, which predict load well but cannot say what share of a prediction is caused by the incentive. We propose a continuous-time neural model of household or building load in which the incentive signal can reach the load through only one narrow, named channel. That channel is a small differential equation with four physically meaningful quantities: how strongly the load responds to the incentive, how much energy can be deferred, how quickly deferred energy is paid back, and what fraction is shed for good. Everything else, including weather, occupancy, and daily routine, is handled by an unconstrained neural component that never sees the incentive.
The bet is that interpretability constraints belong only where a decision depends on them. Grid operators do not need to understand why a house uses 1.3 kW at 7 pm on a Tuesday. They do need a trustworthy, editable account of what an incentive does to that house, including the rebound that synchronized automation can turn into a new peak (He et al. study this herd effect). By confining the constraint to the response channel, we expect to keep nearly all the accuracy of a black-box model while recovering treatment effects with lower bias than both black-box counterfactual rollouts and the linear estimators used in current evaluation practice.
Core Claim
A neural ordinary differential equation (neural ODE) is a model whose hidden state changes continuously in time according to a learned rate-of-change function. It can be queried at any timestamp and handles unevenly spaced observations without first forcing them onto a fixed grid (Chen et al.). We give the model two parts. The first is a latent baseline state $x(t)$ that evolves under an unconstrained neural vector field driven by weather and calendar inputs. The second is a one-dimensional deferred-energy state $d(t)$ that is driven only by the incentive $u(t)$:
$$\frac{dd}{dt} = (1-\rho)\,r(t) - \frac{d(t)}{\tau}, \qquad r(t) = \kappa\, u(t)\, \sigma\!\left(1 - \frac{d(t)}{D}\right),$$
The symbols are defined as follows. $r(t)$ is the instantaneous load reduction. $\kappa$ is the response strength in kW per unit of incentive. $D$ is the deferrable-energy capacity in kWh. $\sigma$ is a smooth saturating gate, so reduction fades as the capacity fills. $\tau$ is the payback time constant in hours. $\rho$ is the fraction of reduced energy that is shed permanently rather than deferred. Observed load is
$$y(t) = g(x(t)) - r(t) + \frac{d(t)}{\tau},$$
where $g$ is a learned readout of the baseline state. The last term is the rebound.
The four quantities $(\kappa, D, \tau, \rho)$ are produced per unit, and per event if needed, by a small encoder of covariates such as outdoor temperature and occupancy. In this sense they form a concept bottleneck. As in Concept Bottleneck Models, which route predictions through human-meaningful concepts that people can edit at test time, the only path from incentive to load passes through named quantities that carry physical units. Unlike Koh et al., these concepts are defined by the structure of the model rather than by annotation, so no concept labels are needed.
Letting the parameters depend on covariates plays the role that the Modified Covariate Method plays in Li and Zhang. Li and Zhang showed that adding covariates additively to a demand response regression inflates the variance of the estimated effect when the baseline is nonlinear and events are rare, and that modeling only the treatment-by-covariate interaction avoids this problem. Our design carries that insight into continuous time. The flexible baseline absorbs the nonlinear physics, and covariates enter the effect only through its interaction with the incentive.
The treatment effect of any incentive schedule is read directly off the model. It is the integral of $-r + d/\tau$ over any window, and it splits exactly into event-window reduction and post-event rebound.
Relation to Prior Work
Three lines of work each supply one piece and leave the combination open.
Structured continuous-time models supply the method. Interpretable Polynomial Neural ODEs restricts the whole vector field to polynomials so that exact equations can be extracted afterward. The authors note that this struggles with saturation and threshold behavior, and those behaviors are exactly what demand response exhibits. Rather than constrain the entire vector field, we constrain only the incentive pathway and let that pathway contain a saturating gate. An et al. calibrate a first-principles PDE-ODE building model and treat internal loads as fixed workday constants, which leaves no room for behavior or price response. That paper constrains our design in a useful way: its stepwise calibration exists precisely because envelope and infiltration parameters become confounded. This warns us that our baseline and response components can confound each other too, so we build the separation into the architecture.
Discrete-time interpretable forecasting motivates the separation. Luo et al. add a concept bottleneck to time-series Transformers, align it to an autoregressive surrogate and hour-of-day, and cut residual connections to prevent "side channel leakage." We borrow the leakage lesson directly: the baseline network never receives $u(t)$. Their model, however, runs on fixed bins, forecasts passively, and has no notion of an intervention.
The demand response literature motivates the problem. Elasticity curves (Zhu et al., Peng et al.) and carbon-aware signaling (Yang et al.) assume a static, block-interval response with no storage dynamics, which is the assumption our deferred-energy state removes. Yildiz et al. show that ensembles of neural ODEs remain stable under irregular sampling in control tasks, but their vector fields are unconstrained. We reuse their ensembling to quantify uncertainty in $(\kappa, D, \tau, \rho)$.
Research Program
We will build the model in JAX with an adaptive solver. Training minimizes load prediction error on data that mixes event and non-event periods, and follows a fixed schedule: the baseline is fit on non-event periods first, and the two components are then fit jointly.
The central scientific question is identifiability. We ask when the structural separation actually recovers the true effect, and how this depends on the event probability $p$, on event duration, and on sampling irregularity. We will study this by varying $p$ in simulation, mirroring the rare-signal regime in which Li and Zhang found linear estimators fragile.
A second question is whether the concepts behave like concepts. Three things should hold. The recovered $\tau$ should track the building's thermal time constant. The recovered $D$ should track the energy stored when the thermostat setpoint is widened. And manually editing $D$, for example to represent occupants tolerating less discomfort after consecutive events, should change predictions in the same way that rerunning the simulator with a narrower setpoint band does. This last check is the analogue of the test-time intervention results in Koh et al.
The third question is the price of the constraint. We measure how much forecasting accuracy on event days is lost relative to an unconstrained model that sees the incentive directly.
Baselines
The naive reference point is a no-response baseline forecaster: our own baseline component trained without any incentive channel. Its error on event days shows how much load variance events actually explain. Any model that fails to beat it on event windows is not learning a response at all.
The strongest existing effect estimator is the Modified Covariate Method of Li and Zhang. It estimates the average treatment effect through treatment-covariate interactions and is robust to nonlinear baselines. In their EnergyPlus experiments they report a normalized estimator variance of 1.020 for this method, 1.000 for simple linear regression, and 3.191 for multiple linear regression with additive covariates. We will run it, together with the simple-regression variant, on identical event assignments and apply it separately to event windows and to post-event windows so that it has a fair chance to capture rebound. Beating it means lower mean squared error in the estimated effect against simulator ground truth at matched $p$, without higher variance, and in particular lower error on the rebound component, which a static binary-treatment estimator cannot represent.
The strongest competitor on accuracy is an unconstrained latent neural ODE (Chen et al.) with $u(t)$ as an input, ensembled as in Yildiz et al.. Its effect estimate is obtained by rolling the model forward with $u$ set to zero and taking the difference. We will also run a bottlenecked Autoformer following Luo et al. on the same data, binned to 15 minutes. That paper reports no per-dataset scalar we can quote, and no Lacuna result exists on Pecan Street forecasting for either model. All neural models receive the same data splits, the same hyperparameter search budget of 40 trials, and the same GPU-hour cap. For the unconstrained ODE, beating it means effect error at least 30% lower while event-day forecast RMSE is within 5% of its value.
The first ablation isolates the key ingredient by letting the baseline network also receive $u(t)$, which reopens the side channel. If effect recovery degrades while forecast error stays flat, then the separation is doing the work. The second ablation removes the deferred state ($\tau \to 0$), which reduces the model to a static elasticity model and tests whether storage dynamics matter.
Why not just…?
Why not just use a black-box forecaster and read off the effect as a counterfactual rollout? This is the most natural competitor, and on regular, dense data it may predict load just as well. The problem is that nothing forces a black-box model to attribute load changes to the incentive rather than to correlated baseline conditions. Events are scheduled on hot afternoons, so heat and incentive arrive together, and the model can explain the dip with either. Li and Zhang's variance inflation under additive covariates is the linear version of this failure. The unconstrained-ODE baseline and the side-channel ablation settle the question directly. If the black-box rollout recovers ground-truth effects as well as our model, the proposal's main claim fails and we will report that.
Why not just use a discrete-time version of the same structure, a linear state-space model with an incentive input? This is a real threat. If telemetry arrives every 15 minutes without gaps, a discrete model with the same baseline-plus-deferral decomposition may match ours, because the structure rather than the continuity may be what matters. Continuity should earn its place in three situations: telemetry is irregular or has dropouts, events start or end partway through a sampling interval, and parameters must transfer across sampling rates. In the last case $\tau$ in hours carries over, while a per-step decay coefficient does not. We will include this discrete twin, train it at 15-minute resolution, and evaluate it at 5- and 60-minute resolution and under 30% random dropout. If it matches us in these conditions too, we will present the contribution as the decomposition rather than the use of an ODE.
Why not just calibrate a first-principles thermal model with a price term? A resistance-capacitance thermal model with an added setpoint-response term is fully interpretable. An et al., however, needed multi-stage calibration to untangle correlated parameters and still reached a next-day demand cv(RMSE) of 17.6% on a single instrumented building, with occupant-driven loads fixed. Hand-built physics is also unavailable for most of the thousands of Pecan Street homes. We include such a model on the EnergyPlus buildings, where its structure is correctly specified, as an upper bound on what physics alone can give.
Proposed Experiments
The primary testbed uses EnergyPlus building simulations, the same simulator family used by Li and Zhang and by An et al., because it supplies ground-truth effects. We run each building twice under identical weather, once with incentive events that widen setpoints or pre-cool and once without, so the true reduction and rebound are known exactly. We will vary event probability from 0.02 to 0.3, event duration from 30 minutes to 4 hours, and sampling from regular 15-minute data to irregular exponential intervals with dropout. The metrics are effect estimation error split into reduction and rebound, estimator variance across 50 seeds normalized as in Li and Zhang, event-day forecast RMSE, and correlation between the recovered $\tau$ and $D$ and their simulator analogs. For the intervention check, we edit $D$ and compare predictions against a simulator run with the matching setpoint change.
The real-data testbed is Pecan Street smart meter data. There, ground truth is unavailable, so we evaluate held-out event-day forecast error, the stability of per-home $(\kappa, D, \tau, \rho)$ across disjoint event subsets, and agreement with the Modified Covariate Method on the event-window average effect. A general forecasting sanity check on the Electricity dataset from the Autoformer suite confirms that the baseline component is competitive when no events occur. As a downstream demonstration, we plug the fitted per-home rebound dynamics into the household-scheduling setting of He et al. and ask whether planning against predicted rebound flattens the post-event peak compared with planning against static elasticities.
The decisive check has three parts, all on EnergyPlus at $p \le 0.1$. The structured model must beat the Modified Covariate Method on total effect error. It must beat the black-box rollout on rebound error. And it must stay within 5% of the black-box model's forecast RMSE. If the side-channel ablation matches the full model on effect error, the separation claim is false. If the discrete twin matches it under irregular sampling, the continuity claim is false.
References
- Fast Electrical Demand Optimization Under Real-Time Pricing
- Neural Ordinary Differential Equations
- Concept Bottleneck Models
- Linear Estimation of Treatment Effects in Demand Response: An Experimental Design Approach
- Interpretable polynomial neural ordinary differential equations
- An inverse PDE-ODE model for studying building energy demand
- Interpretability for Time Series Transformers using A Concept Bottleneck Framework
- Ultra Short-term Load Optimal Dispatching Method for Micro Grid Considering Demand Side Response
- Multi-objective hydropower purchase optimization method for inter-provincial power grids under time-sharing electricity price
- Price-based low-carbon demand response considering the conduction of carbon emission costs in smart grids
- Continuous-time Model-based Reinforcement Learning