PaperlayerView exact revision on arXiv

Source-linked AI summary

Energy Vision--Language--Action: A Controlled Multimodal Benchmark for Intent-Conditioned Residential Energy Management

Lyes Saad Saoud, Oualid Doukhi, Ehsan Reihani, Saeed Sepasi, Deok Jin Lee, Moussa Ayyash, Reza Ghorbani

arXiv:2609.31648v1cs.LGcs.AIcs.SYeess.SY

TL;DR

Residential energy management requires decisions that respond to operating conditions while reflecting user intent, but aggregate error may not show whether a model uses each modality. EVLA constructs a controlled benchmark for intent-conditioned battery-trajectory prediction and finds a strong language-versus-vision asymmetry in its initial study.

  • Problem

    Residential energy decisions must reflect user priorities, and aggregate prediction error may not reveal whether a model responds to language intent or contextual information.

    Source: §1 · p. 1, §1.1 · p. 2

  • Method

    EVLA pairs residential windows with independently varied language objectives and hidden regimes, mapping an RGB energy field, numerical state, and language to a 16-step trajectory generated by a sampling-based reference generator.

    Source: §1.2 · pp. 2–3, §10 · p. 28

  • Results

    Processed language is strongly associated with prediction quality while removing RGB leaves aggregate error essentially unchanged; no-language MSE is 0.8628 ± 0.0001.

    Source: §6.3 · p. 21, §7 · p. 24

  • Takeaways & Limitations

    The results identify whether each modality changes learned decisions, not just whether the benchmark contains multiple modalities, as a central evaluation question.

    Source: §7 · p. 24

    The current controlled release has implementation limitations that constrain conclusions about its target generator, data partitioning, modality interventions, language processing, and reproducibility.

    Source: §8 · p. 25

Abstract

from arXiv · show

Vision-Language-Action (VLA) models are studied mainly in robotics, where visual observations and language instructions are mapped to physical actions. This paper introduces Energy Vision-Language-Action (EVLA), a controlled multimodal benchmark for intent-conditioned residential energy management. EVLA frames battery scheduling as a multimodal trajectory-prediction problem in which an RGB energy-field representation, a numerical operating state, and a natural-language objective are mapped to a 16-step battery-action trajectory generated by a finite-horizon sampling-based reference generator. Source windows are derived from public residential electrical-load data, while electricity price, battery state of charge, indoor temperature, and time of day are generated benchmark metadata. A hidden operating regime is encoded only through energy-field texture, enabling paired visual changes while the explicit numerical state is fixed. Crossing 439,203 retained base windows with three hidden regimes and five language objectives yields 6,588,045 multimodal instances. An initial study evaluates 36 configurations over three training seeds using fixed subsets of 5,000 training, 500 validation, and 500 test instances. In the MobileNet-family comparison, removing processed language increases trajectory mean-squared error from 0.3856 +/- 0.0039 to 0.8628 +/- 0.0001, whereas removing vision yields 0.3843 +/- 0.0013, comparable to the full model. The results show strong asymmetry in modality use: the processed-language pathway is strongly associated with prediction quality, while the current RGB pathway provides no aggregate error advantage. These results characterize the fixed pilot subset and executed protocol rather than full-benchmark training. EVLA provides a controlled setting for studying how semantic intent and latent context influence residential energy-action prediction.

1. Introduction

Residential energy management must balance interacting operating conditions with changing resident priorities. EVLA addresses the need for controlled multimodal decision learning by pairing fixed residential windows with varied language objectives and visually encoded hidden regimes.

1.1. Motivation and Research Gap

  • Existing learning formulations often map numerical state to actions under a fixed objective, while language interfaces typically translate requests into controller parameters rather than learning intent-conditioned trajectories.EVLA’s research gap is a controlled benchmark jointly relating residential state, contextual observations, language objectives, and continuous battery-control trajectories.Source: §1.1 · p. 2

1.2. EVLA Formulation

  • EVLA maps an RGB energy field, numerical residential state, and natural-language objective to a continuous 16-step battery-action trajectory.The hidden regime is encoded through image texture, and the target is a selected sampled trajectory evaluated under action bounds, battery-state limits, smoothness penalties, objective-dependent costs, and regime-dependent evolution.Source: §1.2 · pp. 2–3

1.3. Illustrative Residential Application Scenario

  • For each retained base window, EVLA varies language objective and hidden regime to test whether predictions respond to requested priorities and visual context.The same measured condition can correspond to different trajectories under different resident objectives; the generated targets are explicitly distinguished from physical closed-loop deployment.Source: §1.3 · p. 3

1.4. Contributions

  • EVLA contributes a paired factorial dataset and constraint-aware sampling-based targets, with 439,203 base windows from 19 household files expanded into 6,588,045 multimodal instances.Each base window is expanded across three hidden regimes and five language objectives; the collection is structurally audited.Source: §1.4 · pp. 3–4
  • EVLA evaluates 36 configurations across three seeds on fixed subsets, and processed language is strongly associated with prediction quality while the current visual pathway contributes little to aggregate error.The evaluation uses 5,000 training, 500 validation, and 500 test instances; removing vision leaves aggregate error essentially unchanged.Source: §1.4 · p. 4

2. Related Work

Prior energy-management work emphasizes optimization, forecasting, and control, while language-conditioned interfaces and VLA benchmarks address different parts of the problem. EVLA combines intent-conditioned battery-trajectory prediction with paired multimodal data construction and sampled reference targets.

2.1. Optimization and Learning for Residential Energy Management

  • Optimization and learning-based energy-management methods commonly use fixed objectives, whereas EVLA treats the operating objective as language and predicts continuous battery-action trajectories.The reference trajectories are produced by a constraint-aware sampling-based generator.Source: §2.1 · p. 4

2.2. Human-Interactive and Language-Conditioned Energy Management

  • EVLA places language directly in trajectory prediction and varies it independently of the operating window, unlike interfaces that convert requests into controller settings.This construction permits testing whether changing the stated objective changes the predicted battery-control trajectory.Source: §2.2 · p. 5

2.3. Controlled Multimodal Learning in Cyber-Physical Systems

  • Paired comparisons hold an operating condition fixed while varying inputs, and EVLA combines that construction with ablations and sensitivity diagnostics to examine information content and model response.The distinction matters because aggregate error alone may not reveal whether a model responds to language intent or contextual information.Source: §2.3 · p. 5

2.4. Vision–Language–Action Models

  • EVLA adapts the VLA relationship among observation, semantic intent, and continuous action to residential battery management rather than robotic demonstrations or natural-camera observations.Its paired variation of language intent and visually encoded latent context supports direct tests of modality response.Source: §2.4 · pp. 5–6

3.1. Overview and Dataset Scale

  • EVLA derives 439,203 retained operating windows from 19 REFIT households and expands each across three hidden regimes and five language objectives.The regimes are named low, neutral, and high hidden thermal conditions and encoded as coarse, medium, and fine image texture; the objectives include energy saving, comfort, balanced, grid support, and eco.Source: §3.1 · p. 6

3.2. Operating-Window Extraction and Stored Metadata

  • Each 300-sample residential operating window becomes a 256 × 128 × 3 RGB energy-field image; its channels encode complementary source-window information, while hidden regimes generate regime-specific spatial textures.Source: §3.3 · p. 7

3.4. Controlled Expansion over Regime and Language

  • Every base window yields 15 variants by crossing three hidden regimes with five language objectives; language is assigned independently of the physical operating state.Changing the objective alters the text and target-generation cost, while changing the regime alters visual texture and reference-generator dynamics.Source: §3.4 · p. 8

3.5. Sampling-Based Reference Trajectory Generation

  • A finite-horizon sampling-based generator selects the lowest-cost of 64 candidate action sequences using numerical state, regime, and language objective.The objective determines the stage-cost emphasis, and the resulting targets are benchmark references whose physical interpretation depends on simplified battery and temperature updates.Source: §3.5 · p. 8, §3.6 · p. 8

4. EVLA Trajectory-Prediction Model

EVLA predicts a 16-step battery-action trajectory from an RGB energy field, numerical state, and processed language, with hidden regime information available only through visual texture. Its one-pass model combines learned multimodal routes with a deterministic baseline, while reported errors use an internal action scale rather than physical kW.

4. EVLA Trajectory-Prediction Model

  • The predictor maps an RGB energy field, three-dimensional numerical state, and processed instruction to 16 actions in one pass; hidden regime information must be inferred from visual texture.The regime is not supplied as a numerical input.Source: §4 · pp. 8–9

4.1. Target Representation and Action Scale

  • Reported errors use the evaluator’s internal action scale, not physical kW, and the implementation scale is not a physical bound from the reference generator.The reference candidates have maximum magnitude 1, while P_scale = 5 is an implementation-level learning and evaluation scale.Source: §4.1 · p. 9

4.2. Input Representation and Preprocessing

  • The visual input is resized and normalized per channel, while the evaluated state contains price, SOC, and indoor temperature but excludes hour of day.The executed preprocessing omits canonical ImageNet normalization; per-image, per-channel min–max normalization can suppress absolute channel extrema and inter-image amplitude differences, and no preprocessing ablation was performed.Source: §4.2.1 · pp. 9–10, Figure 3 · p. 10, §4.2.2 · p. 10

4.3. Visual Encoder; 4.4. Language Encoder and Loader Objective Remapping

  • The visual encoder produces 16 projected tokens and a mean embedding, while the simple language pathway remaps instructions to internal objectives before encoding them.The main configuration uses MobileNetV3-Small, initialized from ImageNet weights and frozen except for its final three feature blocks; the evaluated language pathway uses remapped and paraphrased instructions.Source: §4.3 · p. 11, §4.4 · p. 11

4.5. Vision–Language Action-Token Conditioning; 4.6. FiLM Conditioning and Transformer Sequence Model

  • Each learned horizon token receives position-matched visual modulation and the shared language embedding; FiLM supplies additional numerical, global-visual, and language context to the Transformer.The 33-token sequence contains 16 visual tokens, one language token, and 16 conditioned action tokens; future action tokens are masked, but all actions are predicted in one forward pass.Source: Figure 4 · p. 12, §4.5 · p. 12, §4.6 · p. 13

4.7. Residual Decoder and Deterministic Baseline

  • A shared decoder predicts a bounded learned residual that is added to a repeated, non-learned baseline determined by numerical state and inferred objective.The baseline rules are specified in Eq. (31); no final clipping is applied after adding the residual.Source: Figure 6 · p. 14, §4.7 · pp. 14–15

4.8. Composite Training Objective

  • The composite objective includes trajectory regression and regularizers, but peak-objective smoothness weighting depends on batch composition and the SOC proxy uses fixed efficiency 0.95.The first smoothness reference is the first target action, not a measured pre-horizon control; the fixed-efficiency proxy does not reproduce regime-dependent reference-generator efficiencies.Source: §4.2.3 · p. 11, §4.8 · pp. 15–16

4.10. Inference Behavior and Modality Routing

  • At inference, the model predicts all actions without updating residential state between steps; modality ablations measure dependence on implemented inputs, not controlled physical or semantic interventions.The SOC recurrence appears only in the training objective.Source: §4.10 · p. 16

5. Experimental Protocol

The protocol evaluates multimodal trajectory prediction through controlled input ablations, backbone and architecture comparisons, and numerical baselines. Results are reported on a fixed subset, using three training seeds and implementation-defined metrics.

5.2. Dataset Partitioning and Experimental Subset

  • The study uses fixed subsets of 5,000 training, 500 validation, and 500 test instances; the split is not claimed to be base-group-disjoint, overlap-free, or leakage-free.The selected 6,000 instances are approximately 0.091% of the complete collection, so results do not establish performance after full-benchmark training.Source: §5.2 · pp. 17–18

5.3. Experimental Configurations

  • The study evaluates 36 named configurations, including modality settings, visual backbones, architecture variants, and numerical-state-derived baselines.The MLP and LSTM are not input-identical controls: the LSTM also receives previous predicted actions and a normalized time-step signal.Source: §5.3 · p. 18, §5.3.1 · p. 18, §5.3.2 · p. 18

5.4. Training Procedure

  • Learned configurations monitor validation MSE for early stopping, but testing uses the in-memory stopping-epoch model rather than reloading the lowest-validation-MSE checkpoint.The runs use seeds 42, 43, and 44, yielding 108 configuration–seed evaluations.Source: §5.4 · p. 19

5.5. Diagnostic Input Interventions

  • Input diagnostics scramble processed instructions or images, or zero the image; language scrambling leaves the internal objective label unchanged.Disturbance-labelled results are excluded because the loader does not apply a disturbance operator.Source: §5.5 · p. 19

5.6. Trajectory-Error Metrics; 5.7. Operational Diagnostic Metrics

  • Trajectory errors use stored-action units, while operational metrics are implementation-defined diagnostics rather than validated physical or economic quantities.The cycle proxy measures action throughput, not battery degradation or physical cycles; the self-consumption proxy is not the conventional on-site renewable-energy fraction.Source: §5.6 · p. 19, §5.7 · pp. 19–20

5.8. Model Size and Inference Timing; 5.9. Statistical Reporting

  • Inference timings are approximate pipeline diagnostics, and three seeds provide only a limited estimate of training variability without formal hypothesis tests or confidence intervals.No warm-up iterations or explicit GPU synchronization are used.Source: §5.1 · p. 17, §5.8 · p. 20, §5.9 · p. 20

6. Results

The audit supports the generated collection’s record structure, while the controlled ablations show that processed language is associated with lower error than the current RGB pathway. These findings describe a fixed experimental subset and protocol.

6.2. Visual-Backbone Comparison

  • MobileNet has the lowest mean MSE among six full-modality backbones at 0.3829 ± 0.0003, but the lowest-to-highest difference is only 0.0025.The comparison does not show meaningful aggregate-error improvement from increased backbone capacity under this subset and protocol.Source: Table 2 · p. 21, §6.2 · p. 21

6.1. Dataset Audit

  • The audit found 439,203 unique base-state groups among 6,588,045 generated instances, an exact 15-fold expansion consistent with three regimes and five language objectives.It found no malformed records, missing state entries, non-16-step trajectories, or inconsistent battery-SOC and indoor-temperature metadata within groups.Source: §6.1 · p. 21

6.4. Architecture Ablation

  • The deterministic base-only policy has MSE 8.0795, while no-residual, MLP-fusion, and full residual FiLM models report 0.3829 ± 0.0001, 0.3832±0.0010, and 0.3922±0.0111, respectively.The no-residual model has the lowest error among these learned variants but is less sensitive to text scrambling than the full FiLM and MLP-fusion models.Source: Table 4 · p. 22, §6.4 · p. 22

6.3. MobileNet Modality Ablation

  • Removing processed language raises MobileNet-family MSE to 0.8628 ± 0.0001 from 0.3856 ± 0.0039 for the full model, while no-vision MSE is 0.3843 ± 0.0013.The no-language ablation concerns the executed processed-text pathway; the loader also reconstructs an internal objective used by the deterministic baseline.Source: §6.3 · p. 21, Table 3 · p. 22, Figure 7 · p. 23

6.5. Neural Baseline Comparison

  • The designated MobileNet entry reports MSE 0.3829 ± 0.0003, compared with 0.3887 ± 0.0001 for the MLP and 0.3893 ± 0.0004 for the LSTM.These are modest differences, and the baselines are not strictly input-matched.Source: §6.5 · pp. 22–23, Table 5 · p. 23

6.6. Lowest-MSE Named Configurations

  • The ten lowest-MSE named entries include no-vision ResNet at 0.3833±0.0006, Swin at 0.3836 ± 0.0003, and SimCLR at 0.3837 ± 0.0003.The table lists named entries, not a definitive ranking of distinct architectures; equivalent settings can vary because initialization and minibatch sequences differ.Source: §6.6 · pp. 23–24, Table 6 · p. 24

7. Discussion

The discussion finds an asymmetry in the current predictor: processed language is associated with prediction quality, while removing the RGB representation leaves aggregate error essentially unchanged. It emphasizes that modality availability does not ensure modality use, and that model selection must consider conditional responsiveness as well as fit.

7.1. Modality Use versus Modality Availability

  • The hidden regime affects reference actions, yet the learned model gains no aggregate accuracy from the image carrying that regime; information availability does not establish predictor use.The experiments do not identify whether this reflects weak regime effects, difficult textures, limited paired coverage, or the global MSE objective.Source: §7.1 · p. 24

7.2. Processed Language and Objective Semantics

  • The processed-language ablation shows substantial use of that pathway, but does not establish language as necessary for low aggregate error across model families or semantic generalization across the five stored objectives.The loader reconstructs an internal objective and samples a paraphrase; stronger semantic tests require held-out paraphrases and interventions that change text and every objective-conditioned component together.Source: §7.2 · pp. 24–25

7.2. Processed Language and Objective Semantics; 7.3. Architecture and Model Selection

  • Models with similar nominal MSE can differ in text-scrambling sensitivity, so trajectory accuracy and conditional responsiveness are not interchangeable selection criteria.The discussion recommends balancing accuracy, modality sensitivity, computational cost, and seed stability.Source: §7.2 · p. 25, §7.3 · p. 25

7.3. Architecture and Model Selection

  • Simpler learned variants achieve errors comparable to or lower than the full FiLM–Transformer model, while the no-residual model is less sensitive to text scrambling.Greater structural complexity is not automatically rewarded by the current loss.Source: §7.3 · p. 25

7.4. Implications for Residential Energy AI

  • EVLA’s counterfactual organization makes modality dependence inspectable, but adding a modality to the architecture does not guarantee that it affects the learned decision.The proposed next steps include group-aware splits, held-out households and paraphrases, paired regime tests, and a cleaner objective pipeline.Source: §7.4 · p. 25

8. Limitations and Future Work

The current release is intentionally controlled, but implementation and evaluation choices limit the conclusions that can be drawn from it.

8.1. Sampling-Based Reference Generator and Action Semantics

  • The reference target is the optimum among sampled candidates under heuristic dynamics, with manually specified costs, an unvalidated temperature proxy, and no independently validated physical action unit.Operational diagnostics use the stored-action scale; deployment-oriented use would require calibrated models, physical units, and closed-loop feasibility checks.Source: §8.1 · p. 25

8.2. Experimental Scale and Data Partitioning

  • The pilot uses 5,000 training, 500 validation, and 500 test instances, but its split does not ensure base-window grouping or remove overlapping source windows.The authors recommend base-state grouping, a purge interval, exact split manifests, learning curves, and household-held-out evaluation.Source: §8.2 · pp. 25–26

8.3. Visual Representation, Preprocessing, and Regime Intervention

  • Per-image channel normalization can remove cross-image amplitude information, and the experiments do not isolate this effect or the ImageNet-normalization mismatch.The authors propose matched ablations of dataset-level scaling and canonical ImageNet normalization.Source: §8.3 · p. 26
  • The no-vision ablation’s aggregate error cannot establish whether predictions respond correctly to regime changes; paired tests should compare predicted and reference regime differences.The proposed test holds state and language fixed; future generators should use a common candidate-action space across regimes.Source: §8.3 · p. 26

8.4. Language Ontology and Generalization

  • The loader maps five generated objective classes to six internal objectives, may replace instructions with paraphrases, and does not reserve disjoint paraphrases for testing.The authors recommend one objective ontology and testing unseen, human-written, ambiguous, conflicting, and compositional requests.Source: §8.4 · p. 26

8.5. Robustness, Training, and Reproducibility

  • Configured disturbance fields are not applied, and the batch-level “peak” smoothness multiplier also increases the weight for non-peak samples in that batch.The authors recommend traceable perturbation tests and per-sample smoothness weighting or removing the objective-dependent multiplier.Source: §8.5 · p. 26
  • Three seeds, unmatched initializations and data order, and testing without reloading the best-validation checkpoint make fine-grained rankings among near-equal configurations unreliable.The authors recommend matched seeds, checkpoint reloading, more independent runs, and comparison-appropriate uncertainty measures.Source: §8.5 · pp. 26–27

8.6. Generalization Beyond the Current Households

  • The data derive from 19 household files, leaving generalization to unseen households, climates, tariffs, buildings, occupants, and device mixes untested.Future versions should withhold complete households and broaden the operating conditions represented.Source: §8.6 · p. 27

9. Reproducibility and Data Availability

The reproducibility materials document the implemented pipeline and support reconstruction of manuscript tables, but exact reproduction has limits.

9.1. Dataset and Audit Artifacts

  • Dataset audit artifacts record metadata diversity and checks for JSON parsing, state fields, trajectory lengths, within-group consistency, and sampled image availability.The stored metadata include 7,496 distinct state-of-charge values and 6,645 distinct indoor-temperature values.Source: §9.1 · p. 27

9.2. Experimental Reproduction

  • The evaluation records subset-selection seed 42 and model seeds 42, 43, and 44, but exact subset reproduction depends on the complete file layout and lexicographic ordering.The materials do not establish that a standalone manifest of every selected filename is available.Source: §9.2 · p. 27

9.3. Determinism and Reproducibility Limitations

  • Python’s process-dependent hash and unrecorded random-generator states can change derived seeds and generated instances, while timing lacks warm-up and device synchronization.The section also calls for stable hashing, explicit seeds, immutable manifests, checksums, and reloading the validation-selected checkpoint before testing.Source: §9.3 · pp. 27–28

9.4. Data and Code Availability

  • The source REFIT data are public, but the generated dataset and source-code repository are not publicly accessible at the time of the preprint.The supplementary material contains result tables, a dataset-validation summary, and the computing environment; exact subset reproduction also depends on the file collection and layout.Source: §9.2 · p. 27, §9.4 · p. 28

10. Conclusion

EVLA pairs residential windows across objectives and hidden regimes to study battery-trajectory prediction. In the initial study, language removal sharply increased error in the implemented MobileNet-family pathway, while vision removal did not improve aggregate error.

10. Conclusion

  • EVLA pairs each retained residential window across hidden regimes and language objectives, while keeping stored operating-state metadata the same and changing benchmark factors and generated targets by design.Its inputs are a synthetic energy-field observation, numerical state, and language objective; targets are 16-step trajectories from a finite-horizon sampling-based generator.Source: §10 · p. 28
  • The implemented multimodal architecture shows imbalanced modality use: removing processed language causes a large error increase, while removing RGB does not improve aggregate trajectory error.Numerical-only MLP and LSTM baselines remain close to the designated MobileNet result, so the language ablation does not establish universal language necessity; configurations with similar accuracy can also differ in text sensitivity.Source: §10 · p. 28
  • The authors identify modality intervention, group-aware and purged splits, validated action semantics, and stronger reproducibility controls as next steps for reliable conclusions about conditional behavior and generalization.They also call for a consistent language ontology, held-out households and paraphrases, and paired regime tests.Source: §10 · p. 28

S1. Purpose of the Supplementary Material

The supplementary material provides details that complement the main manuscript without introducing a separate dataset version.

S1. Purpose of the Supplementary Material

  • The supplement adds dataset validation, counterfactual organization, reference-trajectory construction, complete ablation tables, and computational-environment details to the manuscript’s benchmark, modeling, protocol, and findings.Source: S1. Purpose of the Supplementary… · p. 31
  • The supplementary material refers to the same EVLA benchmark described in the main manuscript and introduces no separate dataset version.Source: S1. Purpose of the Supplementary… · p. 31

S2. Detailed Dataset Validation

EVLA validates 439,203 retained base operating states expanded into 6,588,045 multimodal samples, with three hidden visual regimes and five language objectives varied around fixed physical states. This paired organization supports controlled diagnostics, but does not itself make the train–validation–test partition leakage-controlled.

S2. Detailed Dataset Validation

  • 439,203 retained base operating states expand into 6,588,045 multimodal samples through three hidden visual regimes and five language objectives.Each base state contributes 15 controlled variants.Source: S2. Detailed Dataset Validation · p. 31
  • Validation confirms that related samples preserve stored physical-state metadata while Z and L vary, and that both factors affect target generation by design.Paired grouping supports diagnostics but does not make the executed train–validation–test partition leakage-controlled.Source: S2. Detailed Dataset Validation · p. 31

S3. Paired-Factor Organization

  • Each base operating state holds numerical state S fixed while independently varying hidden visual regime Z and language objective L.The design produces 15 samples per base state.Source: S3. Paired-Factor Organization · p. 32
  • The paired design separates physical-state dependence from modality dependence: language is sampled independently, and hidden-regime information is available only through the energy-field image.Source: S3. Paired-Factor Organization · p. 32

S4. Supplementary Reference-Trajectory Generator Details

The supplementary material details how EVLA reference trajectories are generated and provides qualitative diagnostics, full ablation records, computing-environment information, and reproducibility-package details. The generated scenarios are controlled prediction targets, not certified closed-loop building control.

S4. Supplementary Reference-Trajectory Generator Details

  • For each (S, Z, L) tuple, the sampling-based generator rolls out candidate action sequences under regime-dependent dynamics and stores the lowest-cost sequence as a 16-step reference trajectory.Costs combine objective-dependent terms, state-of-charge handling, and smoothness; the five objectives encode energy saving, comfort, balanced operation, grid support, and eco operation.Source: S4. Supplementary Reference-Trajectory… · p. 32
  • These scenarios provide controlled sampling-based targets for multimodal trajectory prediction, not certified closed-loop building control.Source: S4. Supplementary Reference-Trajectory… · p. 32

S5. Complementary Qualitative Diagnostics

S6. Complete Ablation Results

  • The supplementary tables provide the full configuration-level and seed-level ablation results underlying the main-manuscript summaries.They exclude superseded draft tables and archived result bundles.Source: S6. Complete Ablation Results · p. 34

S7. Complete Configuration-Level Ablation Results

S8. Complete Seed-Level Results

S9. Computing Environment

S10. Reproducibility Package

  • The archival package specifies manuscript and supplementary sources, data and validation records, ablation results, configurations, split identifiers or hashes, environment details, generation scripts, and license statements.The generated EVLA collection is separate from the public REFIT source data and intended for a dedicated dataset host; the full generated collection is not publicly accessible at the time of the preprint.Source: S10. Reproducibility Package · p. 42
Loading 2609.31648v1…