Calendar regime
id just say workday versus non-day. im not familiar with the term calendar regime.
24 Matching Annotations
- Last 7 days
-
energy-paper.ai-expertteam.mywire.org energy-paper.ai-expertteam.mywire.org
-
-
19.97 and 30.26 kWh, after chronological validation and evaluation on 57 common test days. Its MAE was 43.4% lower than the weekly seasonal-naive reference. XGBoost remained competitive, with RMSE 31.11 kWh,
to put this in context, the total daily electricity consumption of the building could be included here.
-
The weather-source comparison assumes that Rotterdam Airport observations and the selected gridded products adequately represent conditions at the building; it does not resolve hyperlocal weather differences. The products also differ in temporal sampling. The main benchmark retains the supplied weather join, whereas the supplementary comparison uses a consistent endpoint assignment; the original meter interval-label convention is not independently documented (Appendix Section 11.4). Although the IFS values come from a documented historical forecast archive, per-value issue times were not retained, limiting verification of availability at each forecast origin. The measured source sensitivity is consequently a retrospective estimate for the tested configuration and period.
maybe first explain why the Rotterdam airport obversations were used and why they give a fair representation of the building condition. E.g., mentioning how far away it is and anything else they might be relevant.
-
Independent evaluation across further buildings and complete seasonal cycles would strengthen the generality of the findings.
this paragraph lists a bunch of limitations and then says future work is needed. Perhaps include a sentence in between mentioning if the limitations were mitigated in this study or how the results are still useful.
-
common test
is this machine learning jargon? test set?
-
competitive accuracy
how accuracy trade-off exactly?
-
Pretrained forecasting changes the investment required to model a new building. With TimesFM, we supply historical consumption, prepare any covariates, and validate the context and inference settings while keeping network weights fixed. This offers a useful starting point when local model training is inconvenient but historical electricity consumption and suitable computing resources are available. Zero-shot forecasting avoids building-specific weight fitting; selecting a suitable forecasting configuration still requires validation.
nice to mention the performance trade-off (if any) when using a pretrained forecasting model here.
-
The result concerns substitution among weather sources; all three configurations retain weather covariates. Verifying the issue times of operational forecast inputs remains a separate requirement.
i can't follow these sentences.
-
XGBoost’s lower non-workday RMSE may be preferable when large hourly forecast errors are especially consequential. TimesFM 3.0’s slightly lower non-workday MAE may be more relevant when average absolute error is the main objective. An operational preference would require specifying those consequences.
it would be useful to mention how much the errors differ are here and discuss whether they are consequential.
-
earlier THUAS study’s concern
what concern exactly?
-
The practical distinction is therefore between a pretrained model with competitive transfer performance and a comparatively inexpensive model fitted to this building.
id expect a longer discussion here. most of this paragraph is a summary of the results so it would be nice to discuss the implications and interpretations in more depth.
-
The present study extends that work through chronological hyperparameter selection, common test days, daily and weekly seasonal-naive references, residual correction, and pretrained TimesFM models. The research question is whether these additions improve accuracy consistently across days and calendar regimes, and how much local fitting they require. The weather-source and extended-period TimesFM analyses examine the limits of conclusions drawn from the shorter winter test.
similar to the above comment, id keep this information to the method or research question section. In the related work section we can mention what the knowledge gaps are.
-
higher-education
im wondering if referring to university buildings is a better sounding term to use instead of higher-education buildings. In English, THUAS is a type of university after all...
-
higher education
For higher-education buildings,
-
We therefore retain the TimesFM 2.5 target-only model as a version reference and evaluate TimesFM 3.0 both with and without covariates.
id keep information about the method of this paper out of the related work.
-
differences
differences in what?
-
their
there
-
a comparison of statistical, machine-learning, recurrent-neural-network, and pretrained foundation-model configurations for the same higher-education building and forecast days, combined with calendar-regime analysis, paired uncertainty estimates, and a weather-source comparison. We examine whether accuracy gains justify building-specific fitting, whether model rankings change between workdays and non-workdays, and how a pretrained model responds to different weather sources. We select configurations on chronological validation days and retain daily and weekly seasonal-naive references. Calendar and schedule covariates are treated as known at the forecast origin; target-day weather is supplied retrospectively.
this describes what was done but I think we should frame the contributions differently. for example: "This study contributes empirical evidence on the conditions under which increasingly complex forecasting models provide meaningful value for higher-education buildings. By evaluating conventional statistical and machine-learning methods alongside recurrent neural networks and pretrained foundation models within a common building-level setting, we distinguish gains attributable to model complexity from those attributable to building-specific adaptation. The analysis further shows how this comparison depends on calendar regime: workdays and non-workdays may constitute substantively different forecasting settings rather than merely subsets of the same dataset. Finally, by jointly considering predictive uncertainty and alternative weather inputs, the study clarifies the robustness and practical interpretability of pretrained models in building-energy forecasting. This provides a temporally validated basis for selecting forecasting approaches relative to seasonal-naive benchmarks, while distinguishing retrospective weather-informed performance from fully operational forecast performance"
-
Higher-education buildings have operating schedules that change between teaching days, weekends, examinations, and academic holidays.
can we say that university buildings a particularily challenging usecase and therefore, a good test of day-ahead forecasting? i.e., can we generalise the results beyond university buildinggs?
-
motivating improved planning and operation of building energy systems.
id use the first paragraph to explain how improved planning and operation can help before talking about day-ahead forecasting in the following paragraph
-
The building sector has been estimated
Buildings are estimated to account ...
-
.
Im missing a main concluding/take-away statement.
-
higher-education building
i think it's helpful to mention the size of the building, when it was built and rough location.
-
Day-ahead electricity forecasts can support the scheduling of building services, but more complex models also require additional training and maintenance.
suggested revision of the first sentence: Day-ahead electricity forecasts can help building operators schedule services such as heating, ventilation, and cooling in line with anticipated demand, potentially reducing energy use and peak loads. Choosing a forecasting model, however, involves trade-offs between predictive accuracy and the training, maintenance, and computational requirements of more complex approaches.
-