HeatNet: Multifaceted Thermal Systems

• HeatNet is a polysemous term signifying diverse thermal-centric systems, including power–heat co-optimization, district-heating, computer vision, and PDE solvers.
• In energy systems, HeatNet models integrate building thermal dynamics with electrical grids using techniques like SOCP relaxations to optimize DER scheduling and reduce losses.
• In computer vision and numerical analysis, HeatNet architectures leverage multimodal fusion and heat-kernel features to enhance segmentation, pose estimation, and high-dimensional PDE solution accuracy.

HeatNet is a polysemous research term rather than a single standardized system. In the literature considered here, it denotes a coupled thermal–electrical architecture for active distribution grids, a broader district-heating modeling and control viewpoint, a multimodal semantic segmentation model using RGB and thermal images, a 6D pose-estimation pipeline based on heatmap regression, and, in the acronymic form HEATNETs, an explainable random-feature method for high-dimensional parabolic partial differential equations (Panagi et al., 29 Jan 2026, Vertens et al., 2020, Aljosevic et al., 8 May 2026, Georgiou et al., 2 Nov 2025).

1. Principal usages of the term

In the provided literature, the term appears in several technically unrelated but thematically thermal-centered lineages. The table summarizes the major usages before the article turns to each in detail.

Usage	Core object	Representative papers
Thermal–electrical co-optimization	Buildings as thermal storage nodes inside network-constrained OPF	(Panagi et al., 29 Jan 2026)
District-heating network modeling and control	Physics-based planning, operation, retrofit, and control of heat networks	(Hari et al., 2024, Grandits et al., 27 May 2025, Krug et al., 2019, Wack et al., 2024, Sollich et al., 2024, Blizard et al., 2023, Gonzalez-Castellanos et al., 2018)
Heat-demand forecasting and regional data	Weather-driven demand prediction and NUTS-3 heat/PtH datasets	(Xie et al., 2018, Heitkoetter et al., 2019)
Thermal-image semantic segmentation	RGB–thermal day–night domain adaptation for driving scenes	(Vertens et al., 2020)
6D pose estimation	Keypoint heatmap regression with RGB or RGB-D inputs	(Aljosevic et al., 8 May 2026)
High-dimensional PDE solvers	Heat-kernel random-feature neural networks	(Georgiou et al., 2 Nov 2025)

This plurality is itself significant. The term is consistently associated with explicit thermal structure, but the mathematical object named “HeatNet” varies from power-network co-optimization to computer-vision architectures and kernel-based neural solvers.

2. HeatNet as thermal–electrical co-optimization

In the most explicit energy-systems usage, HeatNet denotes a coupled power–heat system in which the electrical network is a radial low-voltage feeder and the “heat network” is implicit in the thermal dynamics of buildings served by air-source heat pumps (Panagi et al., 29 Jan 2026). The studied system is a multi-period, day-ahead network-constrained OPF over a 3-day horizon with 30-minute steps, operating on a 61-bus radial LV feeder behind an 11/0.4 kV, 315 kVA transformer. Flexible DERs are EVs, HPs, and rooftop PV, while time-coupled states include EV state of charge, indoor and envelope temperatures, and transformer thermal aging metrics evaluated ex post (Panagi et al., 29 Jan 2026).

The thermal layer is represented by a calibrated 3R2C grey-box building model with indoor-air temperature $T_{in,t}$, envelope temperature $T_{e,t}$, capacitances $C_{in}$ and $C_e$, resistances $R_{in,e}$, $R_{in,a}$, $R_{e,a}$, solar gains, and heat-pump injection $Q_{hp,t}$ split between indoor and envelope nodes by a factor $f_h \in [0,1]$ (Panagi et al., 29 Jan 2026). The discrete-time dynamics are linear in the states and in $Q_{hp,t}$, which permits direct inclusion in convex OPF formulations. Electrical and thermal domains are coupled by

$T_{e,t}$0

while comfort is enforced by hard bounds

$T_{e,t}$1

This makes HP demand a decision variable rather than a fixed profile, and building thermal inertia becomes an operational flexibility resource (Panagi et al., 29 Jan 2026).

The electrical network is modeled with Branch Flow / DistFlow equations on a radial tree, including nodal active and reactive power balances, voltage-drop equations, current magnitudes, and voltage and current limits (Panagi et al., 29 Jan 2026). The only non-convexity is the current constraint

$T_{e,t}$2

which is relaxed to the SOCP inequality

$T_{e,t}$3

The full problem is therefore a large-scale, time-coupled SOCP with linear thermal dynamics, EV SoC equations, and affine DER limits (Panagi et al., 29 Jan 2026).

A central empirical result is that the SOCP relaxation remains exact in practice even though several classical sufficient conditions are violated by heterogeneous low-voltage cables and by reverse power-flow cases (Panagi et al., 29 Jan 2026). In runtime comparisons across DER penetrations from 0–100%, BIM grows from 1.3 s to 887.7 s, non-convex DistFlow from 0.46 s to 367.3 s, and convex DistFlow (SOCP) from 0.37 s to 0.87 s, remaining sub-second throughout (Panagi et al., 29 Jan 2026). Operationally, coordinated scheduling yields approximately 41% reduction in transformer aging, 54% reduction in total losses, and complete elimination of voltage violations. On a representative day, HP energy decreases from 1227 kWh to 866 kWh, EV energy from 259 kWh to 163 kWh, while average indoor temperature changes from 21.0°C to 20.4°C, still within the comfort band (Panagi et al., 29 Jan 2026).

A common misconception is that this HeatNet is a district-heating hydraulic network. It is not. The paper states explicitly that the “heat network” is implicit: each building is treated as a thermal storage node, and there are no pipes, mass flows, or thermal losses in pipelines (Panagi et al., 29 Jan 2026).

3. HeatNet as district-heating network planning, optimization, and control

A second lineage uses HeatNet as a generic label for district-heating or multi-energy heat-network optimization. Within this literature, the dominant mathematical objects are network flows, thermal transport, capacity planning, and operational control under hydraulic and thermodynamic constraints.

At steady state, district-heating operation can be posed as a Thermal Network Flow Optimization (TNFO) problem on a connected directed graph whose edges represent plants, outgoing steam pipes, return hot-water pipes, loads, and pumps, and whose nodes carry pressures and temperatures (Hari et al., 2024). One representative study models a realistic network with 1 steam plant, 45 loads, 68 outgoing pipes, 68 return pipes, 11 pumps, and 134 junctions. It optimizes plant settings and pump boosts while minimizing undelivered and excess heat subject to plant capacity, load balances, pipe heat losses, steam and water hydraulics, junction mass and energy conservation, and pressure and temperature bounds. The resulting nonlinear, nonconvex continuous problem is solved with IPOPT through JuMP (Hari et al., 2024).

Dynamic district-heating control introduces temporal coupling and storage effects. In a hot-water DHN modeled as a connected directed graph duplicated into supply and return nodes, one study treats plant supply temperatures as controls, node temperatures as states, and fixed mass flows from a separate hydraulic simulation as parameters (Grandits et al., 27 May 2025). On the OpenDHN network in Switzerland, with 150 substations and 2 plants over a 3-day horizon at 15-minute resolution, continuous optimal control yields energy savings of around 8%, cost savings of roughly 12% under dynamic energy pricing, and total runtime of less than 5 minutes on a standard desktop per experiment (Grandits et al., 27 May 2025). The key mechanism is the use of network water volume as distributed thermal storage.

When transient transport and nodal mixing are modeled in higher fidelity, district heating becomes a PDE-constrained nonlinear optimization problem. A complementarity-constrained formulation treats water and heat flow by nonlinear hyperbolic 1D PDEs and models nodal temperature mixing through pooling-like relations recast as complementarity constraints, producing an MPCC later reformulated into a smoother NLP with improved constraint regularity (Krug et al., 2019). The same work develops discretization strategies, penalty formulations, preprocessing for flow-direction simplification, and an instantaneous control method for realistic instances (Krug et al., 2019).

Planning and design studies push the HeatNet idea from operation into topology synthesis. A multi-period topology and design optimization framework for 4th generation district heating combines density-based topology optimization with a nonlinear multi-period physics model (Wack et al., 2024). In the reported case study, moving from worst-case to multi-period design changes the network from separate branched structures to a single integrated meshed topology, increases waste-heat share by 42.8%, reduces project cost by 17.9%, and accommodates producer unavailability at the cost of a 3.1% increase in backup-capacity cost (Wack et al., 2024). For existing networks, a related retrofit framework optimizes producer types, capacities, supplied heat, and supply temperatures across representative periods using GB, HP, ST, and EB technologies (Sollich et al., 2024). As the CO₂ price increases from 0 to 0.3 €/kg, optimal HP capacity rises from 2.5 MW to 8.6 MW, GB capacity falls from 29.5 MW to 23.4 MW, discounted project cost rises from 21.4 M€ to 38.2 M€, and specific emissions fall from 0.145 to 0.080 kgCO₂/kWh (Sollich et al., 2024).

Control-oriented HeatNet models also exist in reduced state-space form. A graph-based automated modeling method for radial DHNs generates a state-space model from topology and component data, validates against laboratory measurements with average normalized RMSE of 0.39 in user mass-flow rates and 0.15 in network return temperature, and shows that a branch-and-bound topology search can reduce network heat losses by 15% relative to a conventional length-minimized topology (Blizard et al., 2023). At the unit-commitment scale, coupled CHP–electricity operation is formulated as a MILP with DC power flow, thermal storage, and detailed CHP performance envelopes; in the reported tests, removing storage increases cost by 1.34%, ignoring network constraints lowers cost by 3.27% but produces infeasible overloads, and decoupled operation increases cost by 11.02% relative to joint operation (Gonzalez-Castellanos et al., 2018).

4. Demand forecasting and regionalized heat data

A third HeatNet-related usage concerns data and forecasting layers that sit upstream of optimization. In a real Swedish district-heating system in Västerås, with about 1.6 TWh/year of supplied heat and more than 14,000 users, heat demand is forecast from the production side using a multi-layer Elman neural network (Xie et al., 2018). The chosen architecture has 8 hidden layers, 15 neurons in each hidden layer, and a 4-hour sliding window. Four datasets are compared: temperature only; temperature plus solar; temperature plus wind; and temperature plus both. On the 2011 test year, ENN-A (temperature only) yields MAPE 6.50%, RMSE 14.7751, MAD 91.6261; ENN-B (temperature + solar) yields MAPE 6.47%, RMSE 14.6969, MAD 71.8131; ENN-C (temperature + wind) yields MAPE 6.43%, RMSE 14.6139, MAD 81.5368; and ENN-D (temperature + solar + wind) yields the best overall performance with MAPE 6.35%, RMSE 14.5358, and MAD 70.6640 (Xie et al., 2018). The paper’s interpretation is precise: wind speed reduces typical error more strongly, whereas solar irradiance is especially effective in reducing worst-case deviations.

For regional and sector-coupled modeling, Germany-wide residential heat demand and power-to-heat capacities have been regionalized at NUTS-3 resolution with 15-minute temporal granularity (Heitkoetter et al., 2019). The dataset uses a special census evaluation to define 729 building categories, derives annual SH and DHW demand per category, classifies installed heating capacity, and estimates electrically covered heat shares from heat pumps and resistive heating technologies. Installed thermal power is approximated by

$T_{e,t}$4

and the outputs are published as open data with DOI 10.5281/zenodo.2650200 (Heitkoetter et al., 2019). This literature does not define HeatNet as a single control architecture, but it provides the regionalized demand and PtH layers needed by HeatNet-style power-to-heat and sector-coupling studies.

5. HeatNet in computer vision

In computer vision, HeatNet names at least two distinct architectures. One is a multimodal semantic segmentation system for autonomous driving that combines RGB and thermal imagery to bridge the day–night domain gap (Vertens et al., 2020). Built on PSPNet with a ResNet-50 backbone, it uses replicated early blocks for RGB and thermal streams, concatenates feature maps after the first two ResNet blocks, and trains with a teacher–student strategy plus output-space adversarial domain adaptation. The associated Freiburg Thermal dataset contains 12,051 daytime and 8,596 nighttime aligned RGB–thermal pairs, for about 20,647 pairs total, together with a labeled test set spanning 13 semantic classes (Vertens et al., 2020). On Freiburg Thermal, the daytime RGB teacher attains 69.4 mIoU by day and 35.7 by night, the thermal teacher reaches 57.0 by night, and the multimodal HeatNet reaches an overall 64.9 mIoU (Vertens et al., 2020). A further contribution is a target-less RGB–thermal calibration procedure based on gradient alignment through a spatial transformer, designed to avoid calibration boards.

A second vision usage applies the name to a modular 6D pose estimation pipeline (Aljosevic et al., 8 May 2026). The system combines YOLOv10m for object detection with a ResNet18-based heatmap regressor that predicts 50 keypoint heatmaps of size 64 × 64 from 256 × 256 RGB crops, followed by PnP RANSAC for pose recovery (Aljosevic et al., 8 May 2026). A cross-fusion RGB-D extension uses dual ResNet18 encoders with residual feature exchange between modalities. On LINEMOD, the best RGB-only model achieves 84.50% mean ADD-based accuracy, while the RGB-D fusion model reaches 92.41% (Aljosevic et al., 8 May 2026). The same study compares keypoint-selection strategies and finds that FPS outperforms CPS in the baseline setting, with mean 6D pose accuracy 86.70% versus 83.34% (Aljosevic et al., 8 May 2026).

These vision usages are unrelated to heat-network infrastructure. In one case the “heat” component is thermal imagery; in the other it is heatmap regression.

6. HEATNETs as heat-kernel random-feature neural networks

The acronymic form HEATNETs designates a family of shallow random-feature neural networks for solving high-dimensional parabolic PDEs (Georgiou et al., 2 Nov 2025). The construction starts from the mild solution representation of a parabolic PDE and uses the Green’s function of the heat operator as the hidden-unit feature map. For i.i.d. random space-time samples $T_{e,t}$5, the approximation takes the form

$T_{e,t}$6

with $T_{e,t}$7, so the network is a single-hidden-layer RFNN whose activations are randomized heat kernels (Georgiou et al., 2 Nov 2025). The paper proves that such HEATNETs are unbiased universal approximators for parabolic PDE solutions, with Monte-Carlo-type convergence rate $T_{e,t}$8 (Georgiou et al., 2 Nov 2025).

To manage singularities and high dimensionality, the method introduces variable transformations and importance Monte Carlo sampling of the integral representation (Georgiou et al., 2 Nov 2025). Training is not performed by deep backpropagation through all weights; instead, with fixed random features, the output weights are obtained from a ridge-regularized least-squares system

$T_{e,t}$9

Benchmark problems extend up to 2,000 dimensions, with reported approximation errors of order 1.0E-05 to 1.0E-07 for problems up to 500 dimensions and of order 1.0E-04 to 1.0E-03 for 1,000 to 2,000 dimensions, using up to 15,000 features (Georgiou et al., 2 Nov 2025). This usage is therefore mathematically and terminologically distinct from both energy-system and vision meanings of HeatNet.

7. Nomenclature, misconceptions, and cross-domain significance

The principal misconception surrounding HeatNet is to assume a single canonical architecture. The literature shows the opposite. In energy systems, the term may refer to building-thermal flexibility embedded in OPF rather than to explicit district-heating hydraulics (Panagi et al., 29 Jan 2026). In district-heating research, it often functions as a generic label for networked heat-system optimization, whether the underlying model is steady-state thermodynamic flow, dynamic optimal control, PDE-constrained MPCC/NLP, topology optimization, or CHP unit commitment (Hari et al., 2024, Grandits et al., 27 May 2025, Krug et al., 2019, Wack et al., 2024, Sollich et al., 2024, Gonzalez-Castellanos et al., 2018). In computer vision, it names architectures that exploit thermal images or heatmap outputs (Vertens et al., 2020, Aljosevic et al., 8 May 2026). In numerical analysis, HEATNETs are heat-kernel random-feature solvers (Georgiou et al., 2 Nov 2025).

A second misconception is that the common label implies methodological similarity. Only a limited structural resemblance exists. In the energy literature, HeatNet typically denotes explicit network constraints, transport, storage, and multi-period coordination. In vision, it denotes multimodal fusion or heatmap regression. In HEATNETs for PDEs, it denotes analytically derived random features. The shared feature is not architecture but the direct encoding of thermal structure into the model class. This suggests that “HeatNet” functions less as a stable proper noun than as a recurrent naming pattern for systems in which heat, thermal imagery, or heat kernels are treated as first-class computational objects.

For researchers, the term is therefore best read contextually. In power and heat-network studies it usually implies sector coupling, thermal transport, or networked heating. In computer vision it usually signals a thermal-modality or heatmap-based network. In numerical PDEs it denotes a specific kernel-based approximation family. Any technical discussion of HeatNet that omits this contextual disambiguation risks conflating fundamentally different mathematical models, optimization problems, and implementation goals.