acceptodds
Under review as a conference paper at ICLR 2027

ALM: Lightweight Adaptation of Foundation Models for Joint Weather-Power Forecasting

Abstract

Weather information is crucial for forecasting power variables such as wind power, photovoltaic power, and load. Existing power forecasting methods typically treat weather variables as exogenous inputs and generate site-level forecasts. However, such localized modeling makes it difficult to capture how continuous weather processes propagate across space, evolve over time, and persistently affect power dynamics. Weather foundation models are well suited to representing these continuous atmospheric processes, yet how to extend their forecasting capability to joint weather-power prediction remains insufficiently explored. To address these limitations, we propose , a lightweight adaptation approach that supports weather-power joint forecasting and bidirectional information exchange between weather and power backbones. ALM consists of three steps: prefusion alignment, latent memory, and postfusion modulation, which support joint improvements in weather and power forecasting while keeping most backbone parameters frozen. We evaluate the method on a gridded weather-power benchmark for the contiguous United States (CONUS) and further validate its forecasts in downstream power applications. Compared with full fine-tuning, ALM reduces mean 72-hour RMSE by 19.9% across 12 reported variables while using only 9.7% trainable parameters. ALM also reaches the shortest per-epoch training time among all compared methods. Our code is available in the Supplementary Material.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Weather Foundation Models, Weather-Power Joint Forecasting, Parameter-Efficient Fine-Tuning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_J5BqOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ALM, a lightweight adaptation framework for joint weather-power forecasting. The method integrates pretrained weather and power models through three components: prefusion alignment, latent memory interaction, and postfusion modulation. It enables bidirectional information exchange while keeping most backbone parameters frozen. Experiments on a constructed CONUS weather-power dataset demonstrate improvements in forecasting accuracy and parameter efficiency, with additional evaluations on downstream power forecasting tasks.

Critical strengths
  1. The paper explores bidirectional information exchange between weather and power forecasting, moving beyond conventional approaches that primarily treat weather as an exogenous input. This is a practically relevant research direction with potential applications in renewable energy forecasting.
  2. The paper compares ALM against a broad range of parameter-efficient fine-tuning methods, including recent approaches such as WeatherPEFT. It also provides component ablations, computational efficiency analysis, downstream applications, and cross-backbone experiments, demonstrating considerable experimental effort.
Critical weaknesses
  1. A central concern is that the current experiments do not sufficiently distinguish whether the improvements arise from incorporating additional power information or from the proposed ALM interaction mechanism itself. Although Table 4 provides component ablations and a power-to-weather injection experiment, these comparisons do not fully isolate the two effects. More controlled comparisons under matched conditions, such as using the same weather-power inputs with simpler fusion mechanisms, would help clarify the independent contribution of ALM.
  2. As described in Appendix A.4, the gridded wind and solar capacity factors are largely constructed using meteorological information and physical conversion models. Therefore, it remains unclear whether the improvements in weather forecasting originate from genuinely complementary power information or from transformed representations of existing weather variables. Further clarification or controlled experiments would strengthen the evidence supporting the claimed bidirectional information exchange.
  3. Although the paper includes extensive comparisons, downstream applications, and cross-backbone experiments, the evaluations primarily rely on the same author-constructed CONUS weather-power dataset. The cross-backbone experiments also use this underlying data source. Additional validation on an independent dataset or a different geographic region would provide stronger evidence of the framework's generalizability.
Questions for authors
  1. Could the authors further clarify how much of the forecasting improvement comes from introducing additional power information versus the ALM interaction mechanism itself? Would comparisons with simpler fusion strategies under matched input conditions help isolate these contributions?
  2. Given that the gridded wind and solar variables are largely derived from meteorological information, how can the authors distinguish genuinely complementary power information from transformed weather representations?
  3. Could the authors further discuss the generalizability of ALM beyond the constructed CONUS benchmark, particularly to independently collected datasets or different geographic regions?
3 comments
AuthorsOpenReview ↗

Responses to Reviewer J5Bq

We thank the reviewer for recognizing that bidirectional exchange moves beyond treating weather as an exogenous input and for noting the breadth of our experiments.

C1. Separating the contribution of power information from the ALM interaction mechanism.

We thank the reviewer for this insightful question. Table 4 already separates the two: closing the power$\to$weather path (η=0\eta=0) worsens all 9 weather variables, an information effect, and removing any ALM component worsens all 12 variables, a mechanism effect. The PEFT comparison is confounded, however, because ALM adds a frozen PowerSwin pretrained on the training split. We therefore compared simpler fusion with the same BitFit Aurora and frozen PowerSwin, inputs, 6→24→726\to24\to72 h schedule, and trainable parameters within ±5%\pm5\% of ALM. The last two fusion rows apply the E1 DoRA or AdaptFormer setup to Aurora and attach PowerSwin through concatenation + MLP (72-h RMSE):

Fusionzz2t2t10u10umsl\mathrm{msl}windsolarload
A-WPFT (single backbone)928.995.343.38636.800.2190.14720.25
Concatenation + MLP1083.754.804.71711.590.2180.18933.13
Bidirectional cross-attention1089.604.633.51768.500.2130.26032.12
Cross-attention, weather$\to$power1014.005.283.97946.340.1990.25831.14
Concat. + MLP + DoRA828.323.893.01665.630.1980.13832.68
Concat. + MLP + AdaptFormer757.803.982.85642.290.2270.16825.32
ALM638.533.312.82554.410.1870.12719.97
  1. Simple dual-backbone fusion is worse than A-WPFT on zz, msl\mathrm{msl}, solar, and load, so adding the power backbone alone does not produce the gain.
  2. With the same backbone, DoRA and AdaptFormer improve over their single-backbone versions in every column shown (Tables 1–2) but stay behind ALM, so the backbone does not explain ALM’s advantage over PEFT baselines.
  3. ALM is best on all 12 variables, so its gain comes from prefusion alignment, latent memory, and postfusion modulation.

We added these baselines to a new subsection of App. D.3.

C2. Complementary power information versus transformed weather representations.

We appreciate this question. Beyond ERA5, the simulated fields encode power-system information (turbine curves, EIA-860 plant data, inverter limits), and load uses no ERA5, so they carry complementary power information (see bpXH-C1). Their ERA5 inputs still include variables outside the nine weather inputs (u100u100, v100v100, d2md2m, spsp, ssrdssrd, fdirfdir). To remove this route, we built observation-constrained wind and PV fields from power-system records only. Each balancing authority’s (BA) hourly EIA-930 generation is split among its plants (EIA-860M) by capacity and the same-month EIA-923 utilization of the preceding year, and the plant outputs in each 0.25∘0.25^\circ cell are divided by the cell’s capacity, with zero where no plant exists. Load stays OEDI. The fields cover 2020–2023, so all four models (and PowerSwin) are retrained on 2020–2021, validated on 2022, and tested on 2023, with wind and solar scored against these fields and missing BA-hours excluded (72-h RMSE):

Methodzz2t2t10u10umsl\mathrm{msl}windsolarload
A-WFT836.756.013.64813.86–––
A-WPFT1080.604.984.14720.110.03150.044322.24
DoRA974.424.923.20503.050.03930.037020.14
ALM554.483.233.18632.420.03090.035314.14
  1. Trained on these fields, all three methods lower the average weather RMSE of A-WFT trained on the same period (on 6, 7, and 8 of the 9 variables), so power information without ERA5 also helps weather forecasting.
  2. ALM is best in six of seven columns (DoRA is best on msl\mathrm{msl}), so its advantage does not depend on simulated targets.

We added this experiment as a new App. D.5.

C3. Generalizability beyond the constructed CONUS benchmark.

Forecasts are already scored against EIA-923 plant and OEDI county data (Table 3), and hourly EIA-930 (bpXH-C2). We added a held-out year (train 2016–2020, validate 2021, test 2022) and metropolitan France (train 2020–2021, validate 2022, test 2023). In France, each method is retrained with ERA5 weather and RTE regional load allocated to 0.25∘0.25^\circ cells by population. Given time limits, all methods use the single-step budget of Table 4 (72-h RMSE):

SettingMethodzz2t2tmsl\mathrm{msl}windload
Held-out yearA-WPFT868.455.95733.820.23118.53
DoRA860.655.76933.590.19937.74
ALM744.733.76716.140.17126.73
FranceA-WPFT1350.088.53536.39–10.05
DoRA896.607.76869.74–15.19
ALM658.236.36784.34–8.66

ALM ranks first on four of five variables in the held-out year and on three of four in France, and lowers the average weather RMSE of A-WFT by 29% and 26%. Its advantage is therefore not tied to the 2023 test year or to CONUS. We added both settings as a new App. D.6.

Related Revisions

  1. App. D.3 (C1).
  2. New App. D.5 (C2).
  3. New App. D.6 (C3).

Thank you for the response. The additional experiments address my main concerns, so I am raising my score from 2 to 3.

AuthorsOpenReview ↗

Responses to Reviewer J5Bq

We thank the reviewer for the careful reassessment and constructive feedback. We are glad that the additional experiments have addressed your main concerns, and we appreciate how your suggestions helped us improve the paper.

Reviewer_885EOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a lightweight method that adapts pretrained weather and power models for joint forecasting. ALM aligns features, exchanges information through shared latent memory, and modulates both branches while keeping most backbone weights frozen. Experiments use a gridded dataset covering the contiguous United States (CONUS) from 2016 to 2023. They evaluate weather and power forecasts, as well as downstream plant- and county-level tasks.

Critical strengths
  1. The paper studies whether power representations can improve weather forecasts. Figure 2 clearly presents the three-scale design. Experiments cover multiple weather variables, power types, and lead times.

  2. The paper provides a broad comparison including full fine tuning, zero shot backbones, and nine PEFT methods. It achieves the best RMSE and ACC on all nine weather groups in Table 1 and on five of six power metrics in Table 2. It also reports trainable parameters and per epoch training time.

  3. The paper ablates its components in Table 4 and Appendix D.3. It ranks first on 19 of 22 metrics on the ClimaX-PowerConv backbone in Tables 13 and 14 and includes downstream plant and county tests.

Critical weaknesses

1.The paper does not validate the realism of its gridded wind and PV targets. Capacity factors rely on a physics based power curve with a fixed 0.855 efficiency and unclear calibration with EIA 930 data. Monthly aggregation of EIA 923 generation in Equations 32 and 33 hides errors within each month.

2.The paper anchors its weather loss to the current trainable branch rather than the pretrained checkpoint, so the teacher can drift, as explained in Section 3.4 and Appendix B.2. It provides no ablation without the anchor or with a fixed pretrained teacher to isolate the role of this loss.

3.The paper uses a single seed, one CONUS domain, and one test year. The cross backbone study keeps the same geographic and temporal distribution, and it provides no parameter matched simple fusion baselines. Robustness and the specific ALM design are therefore not established.

Questions for authors
  1. Could you specify how the hourly wind and solar capacity factor labels are constructed, including the exact roles of EIA 930, EIA 923, physical simulation, and calibration? Could you validate them against independent timestamp level generation data, or report an evaluation without monthly averaging?

  2. Why is the anchor computed from the current trainable branch rather than a fixed pretrained checkpoint? Could you add ablations with no anchoring loss and with a frozen teacher anchor, and quantify how much each option preserves the original weather skill?

  3. Could you report multi seed results and, if available, an additional held out year or region? Could you also compare ALM with parameter matched direct cross attention and unidirectional fusion baselines, and report end to end adaptation time or inference cost alongside per epoch time?

4 comments
AuthorsOpenReview ↗

Responses to Reviewer 885E

We thank the reviewer for the detailed review and for recognizing our broad comparison and best RMSE and ACC on all nine weather groups.

C1. Construction and realism of the wind and PV targets, and hourly evaluation.

We thank the reviewer for this question. Because plant-level records lack the spatial and hourly coverage of the ERA5 grid, the wind and PV targets are physical estimates from conversion models. Wind follows an approach similar to [1]: the NREL ATB power curve, with cut-in, rated, and cut-out limits, is selected from the preceding year's statistics only and applied to density-corrected ERA5 100-m wind. The output is scaled by a fixed aggregate efficiency of 0.855 for energy conversion and array losses and clipped to [0,1][0,1]. Factors around this value are common in wind-power models, e.g., [1,2]. PV uses Perez transposition and PVWatts, weighted by preceding-year EIA-860 fixed/single-/dual-axis shares, with EIA-860 orientations and inverter clipping. Load is OEDI county demand allocated with LandScan/NLCD weights that preserve each county's hourly total, without ERA5. Reanalysis-driven simulation is a standard way to obtain spatially complete hourly renewable fields [3,4]. In label construction, EIA-923 monthly plant generation and EIA-930 hourly BA generation served only to choose these model configurations (2020 for wind, and 2020 for PV with a 2021 robustness check). These years lie in the training split, so the 2022 validation and 2023 test data never enter label construction. We added this procedure to App. A.4 and corrected App. A.2.

To expose the within-month errors that Eqs. (32)–(33) hide, we aggregated the target fields and the 72-h forecasts to balancing authorities with EIA-860 capacity and compared them with hourly EIA-930 generation in 2023. The targets deviate from EIA-930 by 8.89% (wind) and 12.48% (PV) of BA capacity in RMSE, and ALM is the best gridded forecast in five of six columns (see bpXH-C2). County load is scored per timestamp. We added this validation to App. D.2.

C2. Reference of the weather anchor, and ablations with no anchor and a frozen teacher.

The anchor reference is the same weather branch without power or ALM residuals (Eq. 19). It therefore bounds the power-induced change rather than freezing the branch. On CONUS under the single-step budget of Table 4, we removed the anchor or replaced its reference with a frozen pretrained Aurora (72-h RMSE):

Anchorzz2t2tmsl\mathrm{msl}windsolarload
None (λanc=0\lambda_{anc}=0)803.285.16695.860.2280.18124.75
Frozen pretrained Aurora785.374.67689.720.2240.17723.86
Current branch (ALM-6h)709.503.87631.040.2110.16621.45
  1. Removing the anchor raises weather RMSE, so the anchor stabilizes the power$\to$weather correction.
  2. The frozen teacher also gives higher weather and power RMSE than ours, since it penalizes the BitFit updates that adapt Aurora to the joint task.

We added this ablation to Table 4 and App. D.3. We also removed the preservation claim and renamed the term "power-injection regularizer" to state this role.

C3. Multiple seeds, held-out year and region, simple fusion baselines, and cost.

We appreciate this suggestion. We are already retraining ALM and the E1 baselines with three seeds under the full 6→24→726\to24\to72 h schedule and will report the 72-h mean ±\pm std in a later response or the final version.

Under the single-step budget of Table 4, we added a held-out year (train 2016–2020, test 2022) and metropolitan France (train 2020–2021, test 2023, load only). ALM lowers the 72-h weather RMSE of A-WFT by 29% and 26% on average, and is best on seven of nine weather variables in the held-out year and on load in France (see J5Bq-C3 & App. D.6).

We also compared ALM with concatenation and bidirectional or weather$\to$power cross-attention fusion sharing its frozen PowerSwin, inputs, schedule, and ±5%\pm5\% parameter budget. ALM is best on all 12 variables at 72 h, and the unidirectional fusion leaves weather RMSE close to or above A-WPFT (see J5Bq-C1 & App. D.3).

We measured end-to-end training time over all stages, including PowerSwin pretraining for ALM, and inference time per 72-h forecast:

MethodTrainable params (M)End-to-end (GPU-h)Inference / 72 h (s)
A-WPFT1256.4272.40.70
DoRA3.19250.90.98
ALM122.41155.40.77

Even including PowerSwin pretraining, ALM needs 43% less training time than A-WPFT, and only SSF and BitFit are cheaper, by less than 4%. We added this comparison to Sec. 4.4 and a new App. D.7.

[1] Chen R, Hobbs B F, Lu Z, et al. Scientific Data, 2026, 13: 86.

[2] Jourdier B. Advances in Science and Research, 2020, 17: 63-77.

[3] Staffell I, Pfenninger S. Energy, 2016, 114: 1224-1239.

[4] Pfenninger S, Staffell I. Energy, 2016, 114: 1251-1265.

Related Revisions

  1. Apps. A.2, A.4, and D.2 (C1).
  2. Sec. 3.4, Table 4, Apps. B.2 and D.3 (C2).
  3. Sec. 4.4 and Apps. D.3, D.6, and D.7 (C3).
Area_Chair_pun3OpenReview ↗

Reviewer 885E, please respond to the authors' comments

Reviewer 885E, please respond to the authors' comments. Thank you.

Thank you for the detailed response. The additional experiments have addressed some of my concerns. I have also read the other reviewers' comments and your responses, and I am willing to raise my score from 2 to 3.

AuthorsOpenReview ↗

Responses to Reviewer 885E

Thank you for your time and thoughtful feedback. We also appreciate your careful reading of the other discussions. Your suggestions have helped us improve the paper, and we are glad the new experiments have addressed some of your concerns. Please let us know if you have any remaining questions or concerns. We are willing to discuss them further.

Reviewer_bpXHOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper connects pretrained weather and power models through alignment, shared memory, and output modulation to enable bidirectional information exchange, while keeping most backbone parameters frozen. The experiments show that, compared with full fine-tuning, ALM reduces the mean 72-hour RMSE by 19.9% across 12 reported variables while using only 9.7% trainable parameters.

Critical strengths
  1. Adapting pretrained weather models to power forecasting has practical value, and bidirectional information exchange is a worthwhile research direction. The methodology is clearly described.
  2. The evaluation includes comparisons with PEFT and full fine-tuning methods, progressively extended training schedules, and architectural ablations under the same single-step budget as ALM-6h.
  3. The performance and efficiency of the proposed method are also impressive.
Critical weaknesses
  1. Wind/PV fields are derived from meteorological data rather than measured generation. Weather-side gains may therefore reflect additional environmental information. Comparisons with weather-only models using equivalent inputs, shuffled power fields, and real power observations are needed.

  2. Comparing monthly averaged forecasts with EIA-923 data does not establish hourly forecasting skill, particularly for ramps and peaks. Hourly validation is needed, or claims should be limited to monthly consistency. The higher county-load RMSE than full fine-tuning (151.34 vs. 105.47 MW) also warrants discussion.

  3. ALM benefits from an additional pretrained power backbone, unlike some PEFT baselines. Comparisons with parameter-matched dual-backbone fusion methods and consistent training budgets are needed.

  4. The anchoring loss does not guarantee preservation of the original model. Its reference is the current weather branch rather than a frozen pretrained checkpoint, allowing both outputs to drift together. Frozen-reference ablations and out-of-domain retention evaluations are needed.

Questions for authors
  1. The authors should verify the consistency of meteorological inputs and observed power data.
  2. Please reply to the weaknesses listed above.
5 comments
AuthorsOpenReview ↗

Responses to Reviewer bpXH

We thank the reviewer for finding bidirectional exchange worthwhile and for recognizing our clear methodology and ALM's performance and efficiency.

C1. Weather gains may reflect additional environmental information in the derived power fields.

We thank the reviewer for this important question. The power fields are not derived from meteorology alone. They also encode power-system information: NREL turbine curves with cut-in/rated/cut-out limits, EIA-860 plant types, capacities, and orientations, plus inverter efficiency with clipping. Load uses OEDI county demand and no ERA5 (App. A.4). The ERA5 inputs of this conversion outside the nine weather inputs are u100u100, v100v100, d2md2m, spsp, ssrdssrd, and fdirfdir. To separate their effect from that of the power fields and of ALM, we ran three controls on the E1 test set (72-h RMSE). A-WFT + construction variables feeds these six variables to A-WFT through the input path A-WPFT uses for power. The shuffled variants retrain ALM with the whole power sequence, or only load, taken from a random date at the same UTC hour:

Settingzzttqq2t2t10u10umsl\mathrm{msl}
A-WFT (weather only)973.454.780.0009575.383.74933.32
A-WFT + construction variables955.595.510.0008183.953.71741.12
A-WPFT (+ power)928.994.260.001075.343.38636.80
ALM, shuffled power733.863.970.0009994.063.67598.10
ALM638.533.080.0007963.312.82554.41
  1. The construction variables improve A-WFT on five of six variables, yet it stays behind ALM on all six, so ALM's advantage does not come from these extra variables.
  2. A-WPFT beats A-WFT + construction variables on zz, tt, 10u10u, and msl\mathrm{msl}, which means the additional power information in the constructed fields benefits weather forecasting.
  3. With the same inputs, ALM beats A-WPFT in every column, so the remaining gain comes from the ALM design.
  4. Shuffling the power sequence removes 57% of ALM's gain over A-WFT on average (44% when only load is shuffled), so time-aligned power content matters.

We added these controls to a new subsection of App. D.3. With observation-constrained wind and PV fields built from EIA records without meteorology, power still improves weather and ALM remains best (see J5Bq-C2).

C2. Does ALM have hourly skill for ramps and peaks, and why is its county-load RMSE higher than A-WPFT?

County load is already scored at every timestamp (App. B.3). For hourly wind and PV skill, we aggregated the 72-h forecasts (CF ×\times EIA-860 capacity) to balancing authorities (BAs) and compared them with hourly EIA-930 generation at every 2023 target. Errors are in % of BA capacity, ramp is the RMSE of 6-h changes, and peak is the RMSE on top-decile targets:

MethodWind nRMSEWind rampWind peakSolar nRMSESolar rampSolar peak
Target fields (simulated CF)8.898.1413.2212.4815.1919.90
A-WPFT23.1817.7531.7912.8017.2419.83
DoRA20.7316.2232.1213.5018.9819.43
ALM19.2214.8931.6112.4817.4719.19

The target row verifies that the simulated target fields agree with observed hourly generation (8.89%/12.48%). Among the gridded forecasts, ALM is best in five of six columns and trails A-WPFT on solar ramp by only 0.23 points, and with the time-series baselines of App. D.2 it is first or second in every column. The remaining baselines will follow in the final version. We added this validation to App. D.2.

ALM is worse on county load (151.34 vs. 105.47 MW, Table 3) although its gridded load RMSE is lower (19.97 vs. 20.25, Table 2). County load is reconstructed from the gridded forecast through a regularized inverse (App. B.3), which is sensitive to local spatial errors. The county gap therefore arises in this reconstruction step rather than from weaker load skill, as now noted in Sec. 4.2.

C3. Additional power backbone and parameter-matched dual-backbone fusion.

We appreciate this suggestion. In App. D.3, we compare ALM with dual-backbone baselines that share its frozen PowerSwin, inputs, schedule, and trainable-parameter budget (within ±5%\pm5\%). Simple fusion is worse than A-WPFT on zz, msl\mathrm{msl}, solar, and load, and PowerSwin improves DoRA and AdaptFormer, but ALM remains best on all 12 variables. The backbone therefore does not explain its advantage (see J5Bq-C1 & 885E-C3).

C4. The anchoring loss does not guarantee preservation of the original model.

The anchor reference is the weather branch without power or ALM residuals (Eq. 19), so the term bounds the power-induced change, not checkpoint drift. No-anchor and frozen-teacher ablations both give higher error (see 885E-C2). We therefore renamed the term "power-injection regularizer" to state this role and removed the preservation claim.

Related Revisions

  1. App. A.4 (C1).
  2. App. D.3 (C1, C3, C4).
  3. App. D.5 (C1).
  4. Sec. 4.2 and App. D.2 (C2).
  5. Sec. 3.4, Table 4, and App. B.2 (C4).

Thanks for the author's reply; the new results clearly address my main concerns in terms of the additional variables and hourly evaluation. I upgraded the final rating from 2 to 3.

AuthorsOpenReview ↗

Responses to Reviewer bpXH

We sincerely thank the reviewer for the careful reassessment and constructive feedback. We are glad that the additional controls and hourly evaluation have addressed your main concerns. Your suggestions have helped us strengthen the experimental validation and improve the manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Review Discussion and Manuscript Revisions

Dear Program Chairs, Senior Area Chairs, Area Chairs, and Reviewers,

Thank you for overseeing the review. We are grateful to all three reviewers for their careful reading and constructive suggestions. Their feedback helped us distinguish the contribution of the power data from that of ALM's design and prompted us to validate our forecasts against hourly generation records.

We appreciate the strengths highlighted in the original reviews.

  • Reviewer bpXH recognized the practical value of joint forecasting, ALM's accuracy and efficiency, and the matched training budgets in the ablations.
  • Reviewer 885E noted the clear three-scale design in Figure 2, ALM's leading scores across all nine weather groups, and the tests with a second pair of backbones.
  • Reviewer J5Bq highlighted the value of feeding power information back into weather forecasting and the breadth of the baselines and downstream evaluations.

ALM enables bidirectional exchange between pretrained weather and power models through feature alignment, shared latent memory, and output modulation. Compared with full weather–power fine-tuning, it uses 9.7% as many trainable parameters and reduces 72-hour RMSE by 19.9% on average across 12 variables (Tables 1–2 and 21).

The revised PDF highlights changes in blue. In the main text, we clarified the training objective and expanded the discussion of county-load accuracy and training cost. Section 3.4 explains that the power-injection regularizer penalizes the difference between forecasts from the same weather branch with and without power inputs. Table 4 adds the requested comparisons with the penalty removed and with a frozen pretrained reference (Section 4.3). We explain how local grid errors affect reconstructed county loads, where full fine-tuning leads (Section 4.2), and include power-model pretraining in the total training cost (Section 4.4).

The appendix contains seven main additions and clarifications.

  • Appendices A.2 and A.4 explain how we constructed the wind and solar capacity-factor fields and how training-year generation records guided the choice of conversion settings.
  • Appendix D.2 evaluates the targets and forecasts against hourly U.S. balancing-authority generation (EIA-930). ALM ranked first or second in all 24 comparisons, including tests of ramps and peaks (Table 10).
  • Appendix D.3.1 tests six additional ERA5 inputs used to construct wind and solar capacity factors. Adding them to the weather-only baseline lowered its 72-hour RMSE on six of nine weather variables. ALM still had lower RMSE than this augmented baseline on all nine (Table 13).
  • Appendix D.3.2 compares ALM with five fusion baselines matched for backbones, inputs, training schedule, and trainable parameter budget. ALM achieved the lowest 72-hour RMSE on all 12 variables (Table 14).
  • Appendix D.5 repeats the E1 experiment with wind and solar fields constructed from power records alone. ALM improved on weather-only fine-tuning on eight of nine weather variables at 72 hours (Table 18).
  • Appendix D.6 adds a held-out U.S. year (2022) and a French dataset with load as its only power variable. With single-step retraining, ALM reduced 72-hour weather RMSE relative to weather-only fine-tuning by 29% and 26% on average, respectively (Tables 19–20).
  • Appendix D.7 reports training time by stage and inference time per 72-hour forecast. ALM used 43% fewer training GPU-hours than full fine-tuning (Table 21).

In the discussion, bpXH and J5Bq said the new experiments addressed their main concerns. Reviewer 885E said some concerns had been addressed. In response to bpXH's follow-up, we clarified that the E1 loss and metrics include zero-filled load cells outside the contiguous U.S. Excluding them changed only 2 of 66 pairwise orderings for 72-hour load RMSE. ALM still ranked first.

Thank you for considering our responses and the revised manuscript.

Sincerely,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.