Educational Campus Bottom-Up Seasonal Demand Analysis

Data Disclaimer

This case study analyzes actual historical and anonymized electricity-demand data used for demonstration and analytical purposes. The original location, customer segment, dates, and identifying details have been generalized and are presented as Sample Town.

The data is historical, but the technology used to analyze it is current. As of 2026, the study applies QuantRisk’s QR ALPS (AI Load Prediction System) and modern AI/ML techniques to identify demand patterns and develop electricity load forecasts.

Historical observations remain valuable for understanding recurring energy behaviors such as seasonality, weather sensitivity, hourly demand cycles, working-day effects, and consumer-segment characteristics. The fixed dataset also provides a consistent basis for evaluating model performance.

Therefore, this case study combines real-world historical energy behavior with contemporary AI forecasting technology. It is intended to demonstrate forecasting methodology and model performance rather than represent current market conditions. Quantitative results are presented only where supported by the underlying analysis.

1. The Case

Educational Sectors or schools have a distinctive electricity demand profile. Unlike continuously operating commercial facilities, school buildings are strongly influenced by academic schedules, weekday operations, weekends, holidays, and seasonal conditions.

This case study examines the development of an AI-powered electricity load forecasting model for the education sector in a Sample Town. The objective was to forecast hourly school electricity consumption for the final week of each season, using historical demand patterns and relevant external variables.

Rather than treating the entire year as a single homogeneous forecasting problem, the analysis evaluated the behavior of school electricity consumption separately across Spring, Summer, Autumn, and Winter.

The forecasting solution was implemented using QR ALPS (AI Load Prediction System), providing an integrated environment for data preparation, exploratory analysis, feature engineering, machine-learning model configuration, automated optimization, forecasting, and performance evaluation.

The exercise was designed as a historical, one-time forecasting study. The final week of each seasonal period was withheld as the forecast horizon, while earlier observations were used for model development and validation.

2. The Data

The analysis uses a proprietary, historical, anonymized hourly electricity consumption dataset representing school facilities in a Sample Town.

Input Data

The forecasting workflow considered several categories of information:

• Historical school electricity consumption: Hourly load observations used as the primary target series.
• Weather observations: Hourly regional weather information, particularly temperature and relative humidity.
• Solar radiation: DHI, DNI, and GHI variables were assessed as potential predictors.
• Calendar information: Working days, weekends, holidays, and other relevant calendar conditions.
• Time-based variables: Hour of day, day of week, and seasonal context.

Forecast Output

The principal output of the model is an:

Hourly school electricity load forecast

The same type of forecasting framework can also be extended to probabilistic forecasting, allowing expected demand to be accompanied by uncertainty ranges rather than relying only on a single forecast value.

3. Understanding School Electricity Consumption

Before selecting the machine-learning configuration, the data science process focused on understanding how school electricity demand changes throughout the day and across the calendar.

The analysis showed that school electricity consumption is particularly dependent on time of day and school operating schedules.

The four seasonal periods used in the study were:

• Spring: March – May 31
• Summer: June 1 – August 31
• Autumn: September 1 – November 30
• Winter: December 1 – February 28

The last week of each season was reserved as the target forecast period. Earlier comparable periods were used to test candidate model configurations and evaluate errors before applying the selected configuration to the reserved forecast horizon.

This approach allowed the forecasting workflow to distinguish between normal school operating periods and periods with little or no school activity.

3.1. Spring Demand Profile

Spring revealed a pronounced daytime operating pattern.

School electricity consumption began increasing at approximately 5:00 AM and continued rising through the morning before remaining elevated during the main school operating period. Demand began declining around 6:00 PM.

The strongest consumption period was concentrated between approximately 8:00 AM and 11:00 AM.

What the Data Revealed

Several characteristics were particularly relevant:

• A clear morning ramp-up in electricity consumption.
• Sustained demand during school operating hours.
• A pronounced daytime peak.
• Significant dependence on the hour of the day.
• Very low weekend consumption compared with weekdays.

The hourly structure made Hour of Day an important model feature.

Weekday and Weekend Differences

The weekday profiles were relatively consistent, with the main demand period occurring during normal school hours.

Saturday and Sunday behaved very differently.

Consumption on these days was close to the minimum level observed in the dataset, with only limited activity appearing during the morning.

This contrast provided strong evidence that the forecasting model needed to understand the difference between school operating days and weekends.

Modeling Implication

The Spring analysis supported the use of:

• Hour of day
• Weekday
• Calendar information
• Historical load
• Model segmentation

The Model Splitter functionality was therefore considered appropriate for separating materially different consumption profiles.

3.2. Summer Demand Profile

Summer exhibited a broader daytime consumption window.

Electricity demand began increasing around 6:00 AM and remained elevated until approximately 7:00 PM before declining. The main peak period extended from approximately 9:00 AM to 4:00 PM.

Intraday Pattern

The Summer profile demonstrated:

• An early-morning increase in demand.
• Sustained daytime consumption.
• A broad daytime peak.
• Declining demand during the evening.
• Strong dependence on the hour of day.

Calendar Effect

The distinction between operating and non-operating days became especially apparent in the weekday analysis.

Weekdays maintained a recognizable operating profile, with the principal weekday peak around 11:00 AM. Saturday and Sunday displayed a substantially different pattern and much lower overall activity, consistent with the reduced use of school facilities.

Modeling Implication

The Summer analysis reinforced the importance of:

• Hour of day
• Day of week
• Weekend classification
• Calendar effects
• Historical load patterns

These variables provide the model with the context needed to distinguish an active school day from a non-operating day.

3.3. Autumn Demand Profile

Autumn continued to display a strongly structured school-day profile.

Electricity consumption began increasing around 4:00 AM and remained elevated through the school operating period before declining at approximately 13:00 PM.

The primary demand window was concentrated around the morning and midday hours.

Key Characteristics

The analysis identified:

• A morning increase begins around 4:00 AM.
• Elevated consumption throughout the school day.
• Peak activity approximately between 9:00 AM and 11:00 AM.
• Strong hourly dependence.

Weekend Effect

The weekend profile was markedly different from the weekday pattern.

Saturday and Sunday showed substantially lower consumption, consistent with school buildings being largely unoccupied. The weekend data therefore represented a distinct demand regime rather than simply a lower version of the weekday profile.

Modeling Implication

The Autumn results supported combining:

• Hour of day
• Weekday
• Weekend indicators
• Historical load
• Calendar information

within the forecasting configuration.

3.4. Winter Demand Profile

Winter showed another distinct school operating pattern.

Consumption started increasing at approximately 3:00 AM, remained elevated through the daytime operating period, and began declining around 9:00 AM.

The strongest demand occurred during the morning, particularly between approximately 7:00 AM and 9:00 AM.

Weekday Behavior

Weekdays followed a relatively consistent pattern, although Monday and Friday showed somewhat lower overall consumption levels.

• The main weekday peak occurred between 6:00 AM and 9:00 AM.
• Saturday and Sunday showed a substantially lower electricity demand profile.
• This was consistent with the limited use of school buildings during weekends.
• The weekend profile included a lower-demand period extending approximately from 11:00 AM through 6:00 PM.

Modeling Implication

The Winter analysis again demonstrated why the forecasting model needs to understand the relationship between:

• Time of Day + Day of Week + School Operating Schedule
• rather than relying solely on historical load values.

4. External Drivers and Correlation Analysis

Historical electricity consumption provides the foundation for the forecast, but external variables can provide additional information about why demand changes. The data science analysis therefore evaluated weather and environmental variables as potential predictors.

The principal variables examined were:

• Temperature
• Solar radiation
• Relative humidity

The objective was not simply to identify statistically correlated variables, but to determine whether an external time series could provide useful information about a specific component of school electricity demand.

4.1. Analyzing Temperature as a Forecast Driver

Temperature can affect electricity consumption through building heating and cooling requirements.

The historical analysis showed that school electricity consumption had a meaningful relationship with temperature across the seasonal periods.

The relationship also varied throughout the day, meaning that temperature can provide information beyond simply distinguishing summer from winter.

During periods of higher daytime temperatures, the school load profile could also show elevated consumption.

Conclusion

Temperature was selected as an external predictor for the AI forecasting model.

Including temperature enables the model to adjust its expected load based on environmental conditions and potentially improve its ability to capture seasonal and intraday variations.

5. Building the AI School Load Forecasting Model

Following the exploratory and correlation analysis, the forecasting workflow was implemented using QR ALPS.

The platform provides a no-code drag and drop environment in which data-processing components, predictors, model configurations, and optimization functions can be assembled through an interactive forecasting workflow.

The overall process followed:

• Data Integration → Data Preparation → Feature Engineering → Model Configuration → Optimization → Training → Forecasting → Accuracy Analysis
• Separate seasonal configurations were used because the school demand profile changed across the year.

5.1. Preparing the Historical Data

Reliable time-series data is essential for machine-learning forecasting.

The first stage therefore included standard preprocessing functions within QR ALPS.

Gap Filling

The platform supports several approaches for handling missing observations, including:

• Linear interpolation
• Daily-pattern filling
• Weekly-pattern filling

Calendar Processing

Calendar information was especially important for the education sector.
The model needed to distinguish between:

• Normal school days
• Weekends
• Holidays
• Other calendar-specific conditions

This helps prevent the model from treating a non-operating school day as though it were a standard weekday.

5.2. Feature Engineering Strategy

The feature set was derived from the behavioral patterns discovered during data exploration. The forecasting configuration considered three main groups of predictors. Different lag structures and moving-average configurations were evaluated for the seasonal models.

The objective was to combine:

• Recent Demand History + Weekly Recurrence + Calendar Context + Environmental Conditions
• This combination gives the machine-learning models multiple sources of information from which to estimate future school electricity consumption.

5.3. Specialized Model Segmentation

The school dataset contains several fundamentally different operating states.
For example:

• A weekday during school hours
• A weekday overnight
• A Saturday
• A Sunday

do not represent the same demand regime.

The Model Splitter functionality was therefore used to create specialized models rather than forcing a single model to learn all profiles simultaneously.

For the Spring configuration, the case study used three intraday segments:

1. Midnight – 5:00 AM
2. 5:00 AM – 2:00 PM
3. 2:00 PM – end of day

This allows each model to focus on a more consistent portion of the daily demand profile.

The exact segmentation can be adjusted based on the behavior of the specific dataset and validated model performance.

5.4. Machine-Learning Model Selection and Optimization

Several machine-learning approaches were evaluated, including:

• NGBoost
• LightGBM
• XGBoost

These algorithms can model nonlinear relationships between electricity consumption, historical observations, calendar variables, and external drivers.

XGBoost, for example, provides numerous parameters that can affect model behavior. Testing these combinations manually would quickly become impractical.

QR ALPS‘s automated machine-learning functionality can systematically evaluate candidate configurations and fine-tune hyperparameters.

This helps balance model complexity against generalization and reduces reliance on manual trial-and-error configuration.

5.5. Model Training and Forecast Execution

Once the model configuration was finalized, the selected machine-learning models were trained using approximately three months of hourly school electricity consumption data for each seasonal model.

The original case-study configuration reported a model training and forecast execution time of approximately one minute for one of the machine-learning models.

Actual runtime will depend on the dataset size, number of predictors, model complexity, optimization settings, and computing environment.

6. Accuracy Results

After training, the models generated forecasts for the reserved seasonal periods.

QR ALPS provides dashboards that bring forecast information and model-performance indicators together.

Spring

Summer

Autumn

Winter

Note: No accuracy level should be generalized to all schools or education-sector load profiles without independent validation.

7. Extending the Case to Continuous Forecasting

Although this case study uses a frozen historical dataset, the same architecture can be applied to a continuously operating forecasting environment.

A production implementation could connect:

• Educational Sector Consumption Data → Data Hub → Data Quality → Feature Engineering → AI Forecasting → Monitoring → Live Dashboards
• As new electricity consumption observations become available, forecast performance can be monitored against actual demand.
• Weather inputs can also be refreshed as new forecasts become available.
• This creates a continuous forecasting loop rather than a one-time analytical exercise.

8. Scenario Simulation

The same forecasting architecture can also support scenario analysis.

Instead of asking only:

• What is the expected school electricity demand?
• users can explore:
• How might demand change under different future conditions?

Potential scenarios can include:

• Higher or lower temperatures
• Alternative weather conditions
• Different humidity assumptions
• Changes in school operating schedules
• Calendar changes
• Other relevant external-driver assumptions

Scenario simulation can therefore help education-sector organizations assess potential demand under alternative future conditions before those conditions occur. Read more here.

9. Conclusion

The Sample Town educational sector case demonstrates how AI-powered load forecasting can address a particularly distinctive energy-consumption profile.

School electricity demand is strongly structured around academic operating schedules, time of day, weekdays, weekends, holidays, and seasonal conditions. The analysis showed that these patterns should be explicitly represented within the forecasting architecture rather than treated as incidental variations in the historical data.

Temperature and solar radiation provided additional context for understanding the relationship between environmental conditions and daytime school demand. Historical load, calendar variables, and time-based predictors complemented these external signals.

The use of specialized model segmentation allowed the forecasting workflow to distinguish different portions of the daily demand profile, while automated machine-learning optimization provided a systematic approach to selecting model configurations.

The final seasonal accuracy results will be added once the validated figures are available.

More broadly, the case demonstrates how a historical forecasting exercise can form the basis of an operational energy intelligence workflow:

• Data Integration → Data Quality → Behavioral Analysis → Feature Engineering → AI/ML Forecasting → Probabilistic Forecasting → Monitoring → Live Dashboards → Scenario Simulation
• For school districts, education-sector energy managers, and organizations managing portfolios of educational buildings, this approach can provide a foundation for more data-driven electricity planning, demand management, and energy-risk analysis.