STAT 638 (Fall 2026), Lecture 36
Overview
Three project teams present Bayesian analyses: Aditya, Aysegul Buber, and Christopher Browy model 1,417 days of household electricity use; Zhe Xu, Isaac Graham, and Jesse Rutt estimate lithium-ion battery degradation across temperatures; and Sanni Cohn, Halil Iseri, and Vaishnavi Garudkar compare models of building energy loads across 768 designs. Samiran Sinha’s questions highlight model selection, time dependence, hierarchical pooling, predictor scaling, and the need to validate predictive performance and explain modeling choices.
Key takeaways
- For the household electricity project, two seasonal harmonics capture patterns that a single annual sine wave misses, while the estimated linear trend is not distinguishable from zero.
- The household model's Student-t innovations address isolated extreme days but cannot predict sustained absences; its simulations miss a 25-day vacation streak.
- Battery degradation differs by temperature: the 43°C group fades about twice as fast as the 24°C group, while 4°C capacity trajectories are better represented by an offset exponential curve.
- Battery cycle residuals may remain serially dependent after fitting battery-specific curves; Sinha recommends considering AR(1) errors because ignoring dependence can make uncertainty intervals too narrow.
- In the building analysis, joint modeling matters because heating and cooling correlate around 0.98, and Model 3's improved fit to tall-building orientation variation comes with reduced predictive interval coverage.
- Regression coefficients do not provide a fair importance ranking when predictors use different scales; Sinha highlights the need to account for the continuous glazing predictor when comparing effects.
Chapters
- Aditya, Aysegul Buber, and Christopher Browy aggregate nearly four years of French household readings at one-minute intervals into 1,417 daily energy totals.
- Their goals are to assess long-term trend, seasonal and weekday patterns, forecast uncertainty, and unusual consumption periods.
- The mean model uses a linear trend, two seasonal harmonics, and day-of-week effects; the target is square-root transformed, and missing days are not imputed.
- The team uses Student-t innovations to accommodate heavy-tailed daily shocks and an AR(1) error process with its coefficient constrained to the stationary range.
- Weakly informative priors include normal regression coefficients, a half-t scale prior, an exponential degrees-of-freedom prior, and a uniform prior for the AR coefficient.
- A 15-parameter Metropolis-Hastings block sampler uses BFGS initialization and a multivariate normal proposal informed by the approximate posterior covariance.
- All reported R-hat values are below 1.01, effective sample sizes exceed 400, and autocorrelations approach zero within roughly 10 to 20 draws.
- Posterior intervals support weekly and seasonal effects, including both harmonics, while the slightly negative trend is uncertain because its interval includes zero.
- One-day-ahead 95% predictive intervals are about 28 kWh wide and level near 31 kWh by day three; simulated data checks reproduce five of six patterns but miss a 25-day absence.
- Sensitivity tests that halve or double prior spreads barely change estimates; model comparisons favor two harmonics, no linear trend, and the simpler AR(1) over higher-order lag models.
- Bayesian and maximum-likelihood estimates are similar, while Bayesian forecast intervals are only 1 to 3 percent wider; the team reports typical use of 25.7 kWh per day and an additional 4.4 kWh on weekends.
- Christopher Browy says Student-t errors were chosen because extreme shocks would distort a normal-error model, but Sinha asks for a direct normal-versus-t comparison.
- The team used Bayes factors and WAIC for model comparisons; Sinha recommends forecasting-error evaluation using a training period and a held-out time period.
- Aysegul Buber explains that an AR(1) coefficient must have absolute value below one for stationarity and discusses trying AR(2) and AR(4), but not AR(3).
- Aditya explains that a second harmonic allows a less symmetric seasonal curve; Sinha asks the team to justify it visually, in addition to its interval and model-comparison evidence.
- Zhe Xu, Isaac Graham, and Jesse Rutt analyze NASA aging data for 25 lithium-ion batteries and more than 2,000 discharge cycles, using a 1.4 mAh end-of-life threshold.
- They remove the first five ramp-up cycles, scale cycle counts by 100, and fit separate temperature-group models at 43°C, 24°C, and 4°C.
- Capacity decline is modeled linearly at 43°C and 24°C, and with an offset exponential curve at 4°C; battery-specific parameters share group-level distributions.
- The models use Student-t errors and weakly informative priors centered using separate battery-level fits; the 4°C nonlinear model uses tighter prior spreads for stability.
- The 43°C batteries degrade about twice as fast as the 24°C batteries; the 4°C group shows substantial between-battery variation and a lower estimated starting capacity.
- Using BRMS with the NUTS sampler, the team reports R-hat values below 1.01, no divergences, and effective sample sizes above 400.
- Prior-width sensitivity checks preserve the estimated first-100-cycle capacity loss, and posterior predictive simulations place about 97% of readings inside a 95% interval.
- Restricted maximum-likelihood mixed models give similar fade rates, while Bayesian intervals are wider for the small 43°C group; both approaches cover the true remaining life in about 8 of 12 cases.
- A 4°C battery with three near-zero failed runs shows the Student-t fit tracking the main data more tightly than the normal-error fit, which is pulled down and becomes wider.
- Sinha asks why the model treats residual observations as independent across cycles; Zhe Xu acknowledges that an AR(1) structure could capture remaining time dependence and that intervals may be too narrow.
- Jesse Rutt explains the hierarchical structure: battery-level intercepts and slopes come from a shared distribution whose mean is estimated from the data.
- The team notes that it did not compare alternative model forms within each temperature group, identifying that as a future improvement.
- Sanni Cohn, Halil Iseri, and Vaishnavi Garudkar study the UCI simulated building energy-efficiency data, comprising 768 configurations and paired heating and cooling loads.
- The configurations span 12 shapes, two building heights, four orientations, and glazing settings; exploratory analysis finds taller buildings have higher and more variable loads.
- Heating and cooling are strongly associated, with a correlation near 0.98, motivating questions about design effects, joint outcomes, prediction for unseen shapes, and low-energy designs.
- Their baseline model represents design characteristics, shape-specific adjustments, and residual variation that can differ between short and tall buildings.
- Model 2 predicts heating and cooling jointly to represent their dependence and improve uncertainty estimates for total energy use.
- Model 3 lets orientation effects vary with glazing and height, and uses separate shape covariance matrices for short and tall buildings.
- The team uses weakly informative coefficient priors, inverse-gamma variance priors in Model 1, inverse-Wishart covariance priors in Models 2 and 3, and blocked Gibbs sampling.
- They retain 8,000 posterior samples per fit and use posterior predictive checks, prior sensitivity analysis, and leave-one-shape-out validation across all 12 shapes.
- Taller building shapes are associated with about 14.5 additional load units in both outcomes; glazing has a stronger heating effect, with a heating-cooling difference of 0.25 units per 10 percentage points for short buildings and 0.57 for tall buildings.
- For variation in cooling across orientations among tall buildings, Model 3 predicts 1.70 against an observed value of 1.82, while Models 1 and 2 underestimate it more.
- Model 3 has the lowest prediction error but lower interval coverage; Model 1 provides the highest overall coverage, illustrating a fit-versus-uncertainty tradeoff.
- Compact, short, unglazed designs rank among the most efficient; all three models favor unglazed S06, but its probability of being the single best design is only 59% under Model 1.
- Restricted maximum-likelihood mixed-effects models produce coefficient estimates and prediction errors similar to the Bayesian Model 1 results.
- Bayesian prediction intervals account for uncertainty in variance parameters and provide higher heating coverage in the reported comparison.
- Cohn, Iseri, and Garudkar conclude that Bayesian modeling contributes most through explicit uncertainty for effect comparisons and design rankings, rather than large gains in point prediction.
- Their synthesis assigns each model a distinct role: Model 1 is an interpretable baseline, Model 2 captures heating-cooling dependence, and Model 3 represents orientation and shape variability.
- Sinha asks how the five numerical geometry variables are used; Iseri explains that the team instead represents shape through a random effect to generalize beyond the 12 observed shapes.
- Sinha questions whether measured geometry variables should enter the regression individually rather than being grouped into a shape effect.
- Garudkar says feature importance was judged from regression coefficients, and Sinha cautions that coefficients are hard to compare across predictors on different scales, especially continuous glazing.
- Cohn says prediction was not the main objective and mentions Gaussian processes as an alternative; Sinha suggests random forests and neural networks as possible benchmarks before closing the class.
Transcript excerpt
0:02 Sinha, Samiran: Alright, one thing, before you start. Sinha, Samiran: Please let, please, know that this video will be posted online, so it is up to you if you want to keep your video open, or you want to keep it closed. I have no discretion, it is your choice. Hoskere Rajiv, Aditya: Okay, Dr. Sinaya, Sinha, Samiran: So, you may start now, the first group. Sinha, Samiran: part of the group. Sinha, Samiran: Please go ahead. Hoskere Rajiv, Aditya: Okay. Hoskere Rajiv, Aditya: Hi, everyone. We are Christopher, Aysegul and Adityitya, and our project forecasts daily household electricity demand. Our goals are to test for a long-term trend, model seasonal and day-of-week patterns, and track forecast uncertainty while flagging unusual periods. Hoskere Rajiv, Aditya: We use nearly 4 years of 1-minute readings from a home in France, aggregated to total daily kilowatt hours, leaving 1,417 well-observed days, using only calendar dates as predictors.
1:18 Hoskere Rajiv, Aditya: A Bayesian framework lets us carry parameter uncertainty directly into forecasts and probability statements. Buber, Aysegul: So we built our model based on the scientific questions and our EDA. The mean consumption is explained by a linear trend, two harmonics, and day of week effects. We looked at our Box-Cox results, and we decided to do a square root transformation on our target variable. Buber, Aysegul: Also, since this is time series data, we started out with an AR1 error process in our main model. Buber, Aysegul: And then lastly, in… Buber, Aysegul: As you can see in the QQ plot, we have heavy tails, so we use T-Innovations with a single scale. Buber, Aysegul: And then we didn't impute the missing days in our model. Buber, Aysegul: So this is our model based on the assumptions that I just stated. For our likelihood, we use a conditional likelihood, which just means we used observed days in the likelihood calculation.
This is the start of the transcript. Save this video free for the whole transcript, searchable and time-coded.
In this course
Related notes
Summary, takeaways, and chapters were generated by AI from the video's transcript and may contain errors. The video belongs to its creator, Samiran Sinha.