Enterprise Demand Forecasting: Commercial Building Bottom-Up Demand Segmentation

Case Study Data and Technology Context

This case study evaluates electricity-demand forecasting using actual historical and anonymized energy data. To protect proprietary information, the original location, customer profile, dates, and identifying attributes have been generalized and are presented under the name Sample Town.

Historical data is particularly useful for forecasting research because it preserves observable demand behavior under known conditions. Patterns associated with seasonality, weather, operating schedules, time of day, weekdays, weekends, and consumer profiles can be isolated and evaluated without introducing changes to the original dataset.

The data should therefore be distinguished from the technology used to analyze it. The observations are historical, while the AI forecasting methodology is based on technology available in 2026. QuantRisk’s QR ALPS (AI Load Prediction System) is used to apply current machine-learning techniques to the dataset, including data preparation, feature engineering, AI model selection, automated optimization, training, forecasting, and accuracy analysis.

This approach demonstrates how today’s AI capabilities can be applied to historical real-world energy data to evaluate forecasting performance and understand electricity-demand behavior. Reported metrics are included only when supported by the underlying analysis.

1. The Business Challenge

Commercial sectors or more commonly office buildings typically have a more structured electricity consumption profile than many other commercial environments. Demand is closely linked to working hours, occupancy schedules, weekdays, weekends, holidays, and weather conditions.

However, this apparent regularity does not eliminate the need for accurate forecasting. Even relatively stable office demand can vary with seasonal conditions, temperature, operating schedules, and changes between working and non-working days.

For this case study, the objective was to develop an AI-powered hourly electricity load forecasting solution for office facilities in a Sample Town.

The forecasting exercise was designed to predict the final week of each season, using historical demand and external information available before the forecast period. Separate seasonal configurations were evaluated to account for differences in the office consumption profile throughout the year.

The solution was implemented with 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 accuracy evaluation.

The broader objective was to determine how an AI-based forecasting workflow could capture the highly structured behavior of office electricity demand while remaining flexible enough to respond to seasonal and weather-related changes.

2. The Data

The case study uses a proprietary historical dataset containing anonymized hourly office electricity consumption data for Sample Town.

The dataset was frozen for analytical purposes, with the final week of each seasonal period reserved as the forecast target. Earlier observations were used for data exploration, model development, configuration, and validation.

Input Data

The forecasting workflow incorporated:

• Historical office electricity consumption: Hourly load measurements.
• Weather data: Hourly regional observations, including temperature and relative humidity.
• Solar radiation information: DHI, DNI, and GHI variables were evaluated as potential external predictors.
• Calendar data: Working days, weekends, and relevant holidays.
• Time variables: Hour of day, day of week, and seasonal information.

Forecast Output

The primary model output is an:

Hourly office electricity load forecast

The same forecasting architecture can also be extended to produce probabilistic forecasts, allowing users to quantify uncertainty around the expected demand rather than relying exclusively on a single point forecast.

3. Data Exploration and Behavioral Analysis

The first stage of the project was to understand how office electricity consumption behaves before selecting a forecasting architecture.

This exploratory process examined hourly profiles, seasonal changes, weekday/weekend differences, and relationships with external variables.

Office demand showed a strong connection with the working calendar. Unlike continuously operating facilities, electricity consumption was substantially lower during non-working periods, particularly on Sundays.

The historical data was analyzed across four seasonal periods:

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

The final week of each season was reserved as the forecast horizon. Historical periods before the target intervals were used to experiment with model configurations and evaluate forecast errors before applying the finalized configuration to the reserved periods.

Spring Office Demand Profile

Spring provided a clear example of the relationship between office operating hours and electricity demand.

The hourly profile showed demand beginning to increase during the early morning, reaching its main operating-period levels during the working day, and declining later in the afternoon. The strongest demand period extended approximately from 8:00 AM through 4:00 PM.

This behavior is consistent with a commercial facility whose electricity consumption is closely associated with daytime operations.

Key Observations

The Spring analysis identified:

• Demand begins increasing from the early morning.
• Consumption remains elevated throughout the main working period.
• The strongest demand occurs during daytime office hours.
• Load decreases as normal office activity winds down.
• The hourly profile is sufficiently structured that hour-of-day is an important predictive variable.

Weekday and Weekend Behavior

A day-of-week analysis provided additional insight.

The weekday profiles were broadly similar, although their absolute demand levels varied. The principal weekday peak occurred around 3:00 PM. Saturday showed a lower-demand profile with a later peak around 11:00 AM. Sunday consumption was extremely low compared with working days, reflecting the limited operation of offices during the weekend.

Modeling Implications

These findings supported the inclusion of:

• Hour of day
• Day of week
• Working-day classification
• Weekend indicators
• Calendar information

The analysis also supported using the Model Splitter capability to distinguish different demand regimes rather than requiring one model to represent all calendar types.

Summer Office Demand Profile

Summer displayed a similar commercial operating pattern but with changes in the timing and level of electricity consumption.

Demand began increasing during the early morning and remained elevated through the main daytime operating period. The strongest consumption occurred between approximately 9:00 AM and 4:00 PM. This made the relationship between office activity and time of day particularly clear.

Key Observations

The Summer profile indicated:

• Early-morning demand ramp-up.
• Sustained daytime electricity consumption.
• Peak demand concentrated during office operating hours.
• Declining consumption during the later afternoon.
• Strong dependence on the hour of the day.

Weekday vs. Weekend

Weekdays followed a relatively consistent pattern, although the level changed slightly from one day to another.

The weekday peak occurred around 3:00 PM. Saturday showed lower overall consumption, with its main peak shifting toward approximately noon. Sunday followed an even lower-demand pattern, consistent with limited office activity. This distinction between weekdays, Saturday, and Sunday provided a strong rationale for separate model treatment.

Feature Engineering Implications

The Summer results reinforced the use of:

• Hour of day
• Weekday
• Weekend classification
• Calendar features
• Historical load lags

The Model Splitter could separate Monday – Friday, Saturday, and Sunday behavior so that each model learns the appropriate operating profile.

Autumn Office Demand Profile

The Autumn period continued to show a strong relationship between electricity consumption and daytime office activity.

Demand started increasing during the early morning and remained elevated through the central part of the working day. The principal high-demand period extended from approximately 9:00 AM to 4:00 PM, after which electricity consumption began to decline.

Intraday Characteristics

The analysis showed:

• Early-morning demand growth.
• High consumption during daytime operating hours.
• A broad daytime peak.
• Declining demand toward the end of the working day.
• Strong hour-of-day dependence.

The weekday profile showed a primary peak around 2:00 PM. Saturday followed a lower-consumption pattern, with its peak occurring around 10:00 AM. Sunday demand was close to zero for substantial portions of the day, consistent with the office operating schedule.

Modeling Implications

The Autumn findings reinforced the importance of distinguishing:

• Working Days ≠ Saturday ≠ Sunday
• The forecasting architecture therefore used temporal and calendar features together with the Model Splitter to allow different models to specialize in the observed demand regimes.

Winter Office Demand Profile

Winter presented a somewhat different intraday structure while maintaining the same underlying relationship with office operating hours.

The load began increasing around 2:00 AM and continued rising into the daytime period before declining later in the afternoon. The principal daytime demand window extended from approximately 8:00 AM to 3:00 PM. Two noticeable peaks were observed around 9:00 AM and 2:00 PM.

Weekday Behavior

Weekdays displayed broadly consistent profiles, although the overall consumption level varied between days.  Mondays tended to show the lowest weekday levels, followed by Fridays.

The main weekday peaks occurred around:

• 9:00 AM
• 2:00 PM

Weekend Behavior

Saturday and Sunday showed substantially lower demand than working days. The weekend peak was observed around 8:00 AM, although the overall load remained considerably below the normal weekday profile.

This provided further evidence that the calendar and day-of-week variables were essential to the forecasting model.

4. External Feature and Correlation Analysis

Once the main temporal patterns had been established, the next step was to investigate whether external variables could improve the forecast. The analysis examined relationships between office electricity consumption and several environmental time series.

Three categories were particularly relevant:

• Temperature
• Solar radiation
• Relative humidity

The objective was to identify variables that could help the AI model explain changes in consumption level and intraday behavior that historical load patterns alone might not capture.

Temperature

Temperature can influence office electricity demand through heating, cooling, ventilation, and changes in building operating conditions.

The historical analysis showed a positive relationship between temperature and office electricity consumption across the examined seasonal periods.

For example, higher temperatures during warmer periods were associated with higher average demand, while lower temperatures corresponded with lower demand in colder periods.

Conclusion

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

This provides the model with information about expected environmental conditions and helps it adjust the forecast level across seasons and during different periods of the day.

Solar Radiation

Solar radiation variables were also evaluated because office buildings have a strong daytime operating profile.

The analysis considered:

• DHI — Diffuse Horizontal Irradiance
• DNI — Direct Normal Irradiance
• GHI — Global Horizontal Irradiance

Solar radiation provides a useful indicator of the transition between daylight and nighttime and exhibits a daily shape that overlaps with the period when office electricity consumption is most active.

The observed daytime structure suggested that solar radiation could provide additional information about intraday and seasonal changes.

Conclusion

Solar radiation variables were considered useful external predictors for office load forecasting.

They can help the model identify daytime conditions and refine its understanding of seasonal and intraday demand patterns.

Relative Humidity

Relative humidity was also analyzed against the historical office electricity consumption series.

The observed relationship suggested that humidity contained additional information about changes in the consumption profile.

Conclusion

Relative humidity was included as a candidate external predictor.

Its purpose was to help the model adapt to environmental conditions that could affect building energy consumption and complement the information provided by temperature.

The final contribution of the variable should be assessed through model validation and feature-importance analysis rather than correlation alone.

5. The AI Office Load Forecasting Model

With the main behavioral and external drivers identified, the next stage was to construct the machine-learning forecasting workflow.

The implementation used QR ALPS, a cloud-based, no-code AI forecasting platform that brings together data processing, feature engineering, model configuration, optimization, and forecasting.

The overall process was structured as:

Data Preparation → Feature Engineering → Model Design → Automated Optimization → Model Training → Forecast Generation → Accuracy Evaluation

Separate seasonal configurations were maintained because the analysis showed meaningful changes in the office demand profile across the year.

Data Preprocessing

The first modeling stage focused on preparing the historical observations for machine-learning analysis.

Handling Data Gaps

Time-series forecasting systems need reliable and continuous observations.

QR ALPS provides multiple approaches for filling missing values, including:

• Linear interpolation
• Daily-pattern methods
• Weekly-pattern methods

The case-study dataset did not contain material missing-data problems. Nevertheless, gap-filling functionality was included as part of the forecasting workflow to provide resilience against potential data-quality issues.

Calendar Processing

Calendar processing was particularly important for the office profile.

The forecasting workflow classified observations according to:

• Working days
• Weekends
• Holidays
• Other calendar conditions

This enables the model to distinguish a normal Tuesday from a Sunday or holiday instead of assuming that all dates follow the same consumption pattern.

Feature Engineering

The feature set was developed from the findings of the seasonal and correlation analysis.

The models used combinations of:

• Historical electricity demand
• Lagged demand values
• Previous-day observations
• Previous-week observations
• Hour of day
• Day of week
• Working-day indicators
• Weekend classification
• Calendar information
• Temperature
• Solar radiation
• Relative humidity
• Moving-average features where appropriate

Different lag structures were evaluated across the seasonal models.

The purpose was to provide the AI system with both short-term demand memory and context about the current operating and environmental conditions.

Model Splitter

One of the important characteristics of the office dataset was the sharp difference between working-day and weekend consumption.

The Model Splitter was therefore configured to create three distinct forecasting models:

• Monday – Friday
• Saturday
• Sunday

This approach allows each model to specialize in a specific operating regime.

For example, the Monday – Friday model can learn the characteristic daytime operating pattern of offices, while the Sunday model can learn the substantially lower consumption associated with non-operating periods.

This segmentation reduces the need for a single model to reconcile fundamentally different demand profiles.

The source case study describes the same three-way weekday/weekend model structure.

Machine-Learning Model Optimization

Several machine-learning algorithms were considered for the forecasting task, including NGBoost, LightGBM, XGBoost. Each provides different capabilities for modeling nonlinear relationships between historical load, calendar variables, and external predictors.

XGBoost, for example, can be tuned across parameters such as:

• Number of estimators
• Maximum depth
• Minimum child weight
• Gamma
• Subsample
• L1 regularization
• L2 regularization

The number of possible configurations can quickly become very large. Rather than relying on manual trial and error, QR ALPS automated optimization functionality was used to search through candidate configurations and fine-tune model parameters.

This helps address two common modeling risks:

• Underfitting: The model does not capture the important characteristics of office demand.
• Overfitting: The model learns historical noise and performs poorly on unseen periods.

Model Training and Execution

Once the forecasting configuration was finalized, the models were trained using approximately three months of hourly office electricity consumption data for each seasonal model.

The historical case study reported that training one of the selected machine-learning models and executing a forecast took approximately one minute.

Actual execution time in a production environment will vary depending on:

• Data volume
• Number of features
• Number of models
• Optimization search space
• Forecast horizon
• Computing environment

The reported runtime therefore represents the configuration used in this specific case study.

6. Forecast Generation and Accuracy Analysis

The trained models were then used to generate forecasts for the reserved seasonal periods. QR ALPS provides visualization dashboards that combine forecast results with actual values and error measurements.

The evaluation included:

• Actual load
• Forecast load
• Point-by-point forecast errors
• MAE
• MAPE

This combination provides both an aggregate view of model performance and the ability to investigate specific periods where forecasts deviated from actual demand.

Reported Forecast Accuracy

The historical Office case study reported the following error values:

Spring

Summer

Autumn

Winter

The results show variation across seasons, with Spring producing the lowest reported MAPE and Autumn the highest.

This variation highlights an important characteristic of load forecasting: even relatively stable commercial profiles can become more difficult to predict when seasonal conditions or external drivers introduce additional variability.

These results are specific to the historical Sample Town office dataset, forecast horizon, features, model configuration, and evaluation methodology. They should not be presented as guaranteed accuracy levels for other office buildings or markets.

7. Why Office Load Forecasting Benefits From Specialized AI Models

Office buildings often have highly recognizable operating schedules, but that regularity creates its own forecasting requirements.

The model must accurately capture:

• Working-hour demand
• Morning ramp-up
• Daytime operating load
• Afternoon decline
• Weekend shutdown patterns
• Holiday effects
• Seasonal weather effects
• Building cooling and heating requirements

A generic time-series model may identify some of these patterns, but a more flexible AI workflow can explicitly incorporate the calendar, environmental variables, and specialized model segmentation.

The Sample Town case demonstrated that weekday, Saturday, and Sunday profiles were sufficiently different to justify separate model treatment.

This is a key reason why configurable AI forecasting architectures can be valuable for commercial energy applications.

8. From Historical Case Study to Operational Forecasting

Although this case study evaluates a frozen historical dataset, the same architecture can be extended to a continuously operating forecasting environment. Read more here.

A production implementation can connect:

Data Sources → Data Hub → Data Quality → Feature Engineering → AI Forecasting → Monitoring → Interactive Dashboards

New electricity consumption observations can be continuously incorporated into the forecasting workflow, while weather and calendar data can be refreshed as new forecasts become available.

This creates the foundation for a more dynamic energy forecasting process rather than a one-time analytical exercise.

9. Probabilistic Forecasting and Risk-Aware Planning

Point forecasts provide an expected demand value, but energy operations often require an understanding of uncertainty.

The forecasting architecture can therefore be extended to probabilistic load forecasting, producing prediction intervals or distributions around the expected electricity demand.

For an energy manager or utility, this can answer questions such as:

• What demand level is most likely?
• How high could demand reasonably become?
• Which hours carry the greatest forecast uncertainty?
• How much operational flexibility may be required?

This additional layer can support more informed energy procurement, capacity planning, demand management, and risk assessment.

10. Scenario Simulation

The forecasting workflow can also support future scenario analysis.

For example, users can examine how expected office electricity demand may change under alternative environmental or operational conditions.

Potential scenarios include:

• Higher or lower temperatures
• Different humidity conditions
• Changes in expected weather
• Alternative calendar assumptions
• Changes in operating schedules
• Other relevant external drivers

Scenario simulation allows organizations to move beyond a single baseline prediction and examine a range of possible future demand conditions.

This is particularly useful when energy decisions need to be made before the actual operating conditions are known. Read more here.

11. Conclusion

The Sample Town office case demonstrates how AI-powered load forecasting can be applied to a commercial electricity profile characterized by strong working-day patterns, low weekend consumption, clear daytime operating peaks, and seasonal weather effects.

The analysis showed that time-of-day and calendar information were fundamental to understanding office electricity demand. External predictors such as temperature, solar radiation, and humidity provided additional context that could help the AI models adapt to changing environmental conditions.

The forecasting architecture combined specialized model segmentation with automated machine-learning optimization, enabling separate models to learn the distinctive behavior of weekdays, Saturdays, and Sundays.