Bottom-up Demand Forecasting by Consumer Type Segmentation: A Seasonal Study on Hospitality Sectors

About the Data and AI Methodology The analysis presented here is based on historical, anonymized energy-consumption data obtained from a real-world case and adapted as Sample Town to protect identifying information. Specific dates, locations, customer details, and other proprietary characteristics have been generalized.

The age of the dataset is intentional. Preserving a historical dataset without changing its underlying observations provides a stable and reproducible environment for testing forecasting methods. Electricity-demand characteristics such as hourly cycles, seasonal variation, weather relationships, and working-day effects can therefore be analyzed consistently.

While the data represents a historical operating period, the forecasting technology applied to it is current as of 2026. The analysis uses QuantRisk’s QR ALPS (AI Load Prediction System) and its contemporary AI/ML forecasting capabilities, including feature engineering, model selection, hyperparameter optimization, model training, and performance evaluation.

Accordingly, this case study should be understood as a demonstration of modern AI applied to real-world historical energy behavior, rather than as a representation of today’s electricity market conditions. All quantitative performance figures are based on the underlying case-study analysis.

1. The Business Challenge

Hotels have distinctive electricity consumption profiles. Unlike many commercial facilities, hotel demand can remain active throughout the day and night, while changing significantly according to guest activity, operating schedules, occupancy-related behavior, weekends, holidays, weather, and seasonal conditions.

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

The forecasting exercise focused on predicting the final week of each season, using historical observations available before the forecast period. Separate seasonal models were developed so that the forecasting system could account for differences in hotel electricity consumption throughout the year.

The project was implemented using QR ALPS, providing an integrated environment for data preparation, exploratory analysis, feature engineering, AI model configuration, automated optimization, training, forecasting, and performance analysis.

Rather than treating hotel demand as a simple repeating time series, the approach examined how consumption changes according to:

• Time of day
• Day of week
• Weekend behavior
• Seasonal conditions
• Temperature
• Humidity
• Historical demand patterns
• Calendar effects

The objective was to develop a forecasting workflow capable of learning these patterns and producing more informed hourly demand predictions.

2. The Data

The analysis uses a proprietary historical dataset containing anonymized hourly electricity consumption data for hotel facilities in Sample Town.

For the purposes of the case study, the historical dataset was frozen so that the forecasting problem could be evaluated consistently. The final week of each season was treated as the target forecasting interval, while earlier observations were available for exploratory analysis, model development, and validation.

Input Data

The forecasting workflow combined several sources of information:

• Historical hotel electricity consumption: Hourly load observations.
• Weather information: Hourly regional weather data, including temperature and relative humidity.
• Calendar information: Holidays and calendar classifications.
• Time-related variables: Hour of day, day of week, weekday/weekend, and seasonal information.

Forecast Output

The primary output of the system is an:

Hourly electricity load forecast

The forecasting architecture can also be extended to generate probabilistic forecasts and uncertainty intervals, providing decision-makers with information about the range of potential future demand rather than only a single expected value.

3. Understanding Hotel Electricity Consumption

Before configuring the AI models, QR data science team examined the historical consumption series to identify recurring patterns and unusual behavior.

This exploratory stage is particularly important for hotel demand forecasting because hotel electricity usage can differ substantially from residential or conventional office profiles.

A hotel may have continuous operations involving guest rooms, restaurants, kitchens, common areas, lighting, heating and cooling systems, laundry, elevators, and other services. Consequently, the shape of the demand curve can change significantly during different periods of the day and across seasons.

The historical analysis showed noticeable seasonal differences in the Sample Town hotel consumption profile, with higher average electricity demand during the Summer period.

For the analysis, the year was divided into 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 forecasting target.

To develop and validate the models, earlier seasonal periods were used to experiment with model configurations and evaluate forecasting errors. Once the appropriate configurations had been established, the models were applied to the reserved forecast periods.

3.1. Spring Demand Profile

The first seasonal analysis examined hotel electricity consumption during Spring.

The hourly demand profile revealed a distinctive pattern that differed from the residential profile examined in other forecasting cases.

Intraday Pattern

The Spring data showed a relatively high level of electricity consumption during the evening, with a notable peak around 10:00 PM.

Additional smaller increases appeared during the morning and late-morning periods, while the lowest consumption levels occurred around the middle of the afternoon.

The observed profile included:

• A major evening peak around 10:00 PM.
• Smaller morning peaks around 6:00 AM.
• Another increase around 10:00 AM.
• Lower consumption during the afternoon.
• A particularly low period around 3:00 – 4:00 PM.

These observations demonstrated that the hotel’s demand profile was strongly dependent on time of day.

Feature Engineering Implication

The analysis therefore supported the inclusion of:

Hour of Day

as an important predictor in the AI forecasting model.

This feature allows the model to learn recurring intraday patterns instead of treating each hourly observation independently.

Weekday and Weekend Patterns

The hourly profile was then analyzed separately for individual days of the week.

The results showed that weekdays followed broadly similar shapes but differed in overall demand levels.

Several patterns were particularly relevant:

• Monday and Friday tended to have lower overall demand than several other weekdays.
• Weekday demand showed morning and evening peaks.
• The evening peak occurred around 10:00 PM.
• Weekend demand followed a different pattern from weekdays.
• Weekend morning activity shifted toward approximately 10:00 – 11:00 AM.
• Sunday evening and Monday morning represented particularly low-demand periods.

These differences indicate that hotel demand cannot be accurately represented using hour-of-day information alone.

Model Configuration Implication

The findings supported the use of:

Hour of day

Weekday / day of week

Weekend indicators

Calendar information

The Model Splitter capability in QR ALPS was also considered appropriate because it allows the forecasting workflow to create separate models for different behavioral groups.

For this case, the configuration separated:

• Weekdays
• Saturdays
• Sundays

This enables each model to learn the characteristics of its own demand regime.

3.2. Summer Demand Profile

The Summer period displayed a different demand structure from Spring.

The historical hourly averages indicated that hotel electricity consumption was generally higher during the evening than during the morning.

The strongest average peak occurred around 10:00 PM, while another smaller peak appeared around midday.

Intraday Characteristics

The Summer profile showed:

• Higher evening demand.
• A dominant peak around 10:00 PM.
• A secondary increase around noon.
• Clear dependence on the hour of day.

The seasonal change in demand reinforced the importance of using a forecasting model that can adapt its expected load level according to both time and seasonal conditions.

Weekday vs. Weekend

The day-of-week analysis showed that Monday through Friday followed similar general shapes, although their overall demand levels differed.

The main observations included:

• A weekday morning peak around 11:00 AM.
• A stronger evening peak around 10:00 PM.
• Weekend morning activity shifted toward approximately noon.

Monday mornings and Sunday nights continued to represent lower-demand periods.

Feature Engineering

The Summer results supported continued use of:

• Hour of day
• Day of week
• Weekend classification
• Calendar features
• Historical load variables

The Model Splitter could again be used to distinguish weekday demand from Saturday and Sunday profiles.

3.3. Autumn Demand Profile

The Autumn analysis revealed another variation in the hotel’s electricity consumption curve.

Average demand remained relatively elevated through the morning and approached midday before declining toward the afternoon minimum.

Two smaller morning peaks were visible, approximately around 7:00 AM and 11:00 AM.

Following the midday period, consumption declined and reached its lower levels during approximately 3:00 – 4:00 PM.

Demand then increased again during the evening, with the primary nighttime peak occurring around 9:00 – 10:00 PM.

Key Characteristics

The Autumn profile therefore contained several distinct intraday phases:

• Early-morning increase.
• Morning peak.
• Late-morning increase.
• Midday decline.
• Afternoon minimum.
• Evening recovery.
• Nighttime peak.

This type of structure demonstrates why an hourly forecasting model needs to recognize multiple intraday regimes.

Day-of-Week Analysis

The weekday analysis indicated broadly similar patterns but with differences in overall consumption levels.

Monday showed comparatively lower demand, with additional variation observed toward the end of the working week.

Weekdays displayed morning and evening peaks, while weekend demand followed a different pattern with the main morning increase occurring around noon.

Sunday evening and Monday morning again appeared among the lower-demand periods.

Modeling Implications

The Autumn analysis supported the continued use of:

• Hour-of-day features
• Day-of-week variables
• Weekend indicators
• Historical demand lags
• Calendar information
• Weather predictors

The Model Splitter configuration could separate weekday and weekend profiles to allow specialized models to capture these differences.

3.4. Winter Demand Profile

Winter displayed the most noticeable change in the hotel’s daily demand profile.

Unlike the other seasons, demand declined progressively from the early morning toward the afternoon.

The lowest levels occurred around 3:00 PM, after which consumption began increasing again.

The demand curve subsequently approached the early-morning level around the evening period, with the strongest nighttime peak occurring around 10:00 PM.

Intraday Characteristics

The Winter profile showed:

• Morning demand around 7:00 AM.
• A gradual decline through the daytime.
• An afternoon minimum around 3:00 PM.
• Increasing consumption from the afternoon onward.
• A significant evening peak around 10:00 PM.

This seasonal change demonstrated why the forecasting system needs to learn the relationship between seasonality and intraday demand behavior.

Weekday and Weekend Analysis

The Winter data showed relatively similar overall behavior between weekdays and weekends, although consumption levels varied.

Saturday showed somewhat higher demand during portions of the middle of the day.

Morning demand generally peaked around 7:00 AM, while nighttime demand peaked around 9:00–10:00 PM depending on the day.

As observed in the other seasons, Sunday evening and Monday morning were among the lower-demand periods.

Feature Engineering

The Winter analysis confirmed the value of:

• Hour of day
• Day of week
• Weekend information
• Historical load patterns
• Calendar variables
• Weather information

The seasonal analysis therefore provided the basis for configuring a dedicated Winter forecasting model rather than assuming that the same demand curve would apply throughout the year.

4. External Feature and Correlation Analysis

Historical demand provides the foundation for load forecasting, but hotel electricity consumption can also be influenced by external environmental conditions.

The next stage of the analysis therefore investigated the relationship between hotel electricity demand and available weather time series.

The objective was to identify variables that could provide additional predictive information about changes in demand levels across seasons and hours of the day.

The analysis focused primarily on:

• Temperature
• Relative humidity

These variables were evaluated against the historical hotel consumption series to determine whether they contained useful information for forecasting.

4.1. Temperature as a Forecast Predictor

Temperature can influence hotel electricity demand through heating, cooling, ventilation, and changes in guest and facility behavior.

The analysis identified a positive relationship between temperature and average hotel electricity consumption during several periods.

Higher temperatures, particularly during Summer and early Autumn, were associated with higher average electricity consumption in the historical data.

Conversely, lower temperatures during the Winter period corresponded with lower average demand in the analyzed dataset.

Conclusion

Temperature should be considered as an external predictor in the AI forecasting model.

Including temperature allows the model to adjust its expected demand level according to environmental conditions rather than relying solely on historical load patterns.

This can be especially useful for forecasting periods where weather conditions differ from the historical average.

4.2. Humidity as a Forecast Predictor

Relative humidity was also evaluated as a potential explanatory variable.

The analysis indicated an inverse relationship between humidity and hotel electricity consumption during the examined historical periods.

This suggests that humidity may contain additional information about changes in the demand profile that is not fully captured by time-based variables alone.

Conclusion

Relative humidity should be evaluated as an external feature in the AI forecasting model.

Its contribution can then be assessed through model optimization and out-of-sample validation to determine whether it improves forecasting performance.

The final feature set should be determined by measured predictive value rather than correlation alone.

5. Building the AI Hotel Load Forecasting Model

Following the exploratory and correlation analyses, the next stage was to configure the AI forecasting solution for the hotel demand profile.

The implementation used QR ALPS, a no-code cloud-based forecasting platform that enables users to assemble data-processing, feature-engineering, and machine-learning components through an interactive interface.

Separate seasonal models were developed because the analysis demonstrated that hotel electricity demand changes meaningfully across the year.

5.1. Data Preparation

Before training the forecasting models, the historical data was prepared for machine-learning analysis.

Gap Filling

Energy datasets can contain missing observations caused by data collection, communication, or other operational issues.

QR ALPS provides multiple approaches for handling gaps.

The historical dataset used in this case did not contain significant missing observations. However, gap-filling functionality was incorporated into the forecasting workflow so that the process could accommodate potential data-quality issues.

Calendar Processing

Calendar information was used to distinguish between:

• Weekdays
• Saturdays
• Sundays
• Holidays
• Other calendar-specific conditions

This is important for hotel demand forecasting because guest activity and hotel operations may differ significantly across calendar types.

Calendar processing therefore provides the AI model with additional context about the type of day being forecast.

5.2. Feature Engineering

The feature configuration was derived directly from the seasonal and external-variable analysis.

Depending on the season, the models used combinations of:

• Historical electricity consumption
• Previous-day demand
• Previous-week demand
• Hour of day
• Day of week
• Weekend classification
• Temperature
• Relative humidity

Different lag configurations were evaluated for each seasonal model.

Load Forecast Model for Hotel Consumer Type for Spring

This allows the model-development process to determine which historical demand observations provide the most useful information for predicting the target period.

The aim was to create a feature set that captures both recurring demand behavior and external drivers of consumption.

5.3. Model Splitter

The hotel demand profile showed meaningful differences between weekdays, Saturdays, and Sundays.

To capture these differences, the Model Splitter capability was configured to divide the forecasting problem into separate behavioral groups.

The model architecture therefore consisted of three forecasting models:

• Monday – Friday
• Saturday
• Sunday

Each model was trained to represent the characteristics of its corresponding day type.

This approach allows the forecasting system to specialize instead of forcing a single model to learn substantially different demand profiles simultaneously.

5.4. Automated Machine-Learning Optimization

Several machine-learning algorithms were evaluated for the forecasting task, each model offers different approaches to learning nonlinear relationships between historical demand, temporal variables, and external predictors.

XGBoost, for example, can be configured using several hyper parameters. Manually testing combinations of these parameters can quickly become impractical.

QR ALPS provides automated model optimization to search through candidate configurations and identify suitable combinations based on the selected validation methodology. This reduces the amount of manual experimentation required.

5.5. Model Training and Execution

Once the model configurations were selected, the forecasting models were trained using approximately three months of hourly hotel electricity consumption data for each seasonal model.

Training a selected machine-learning model and executing the forecast took approximately one minute under the tested configuration. Actual production runtime will depend on different factors. The reported execution time should therefore be interpreted as a characteristic of this particular case-study setup rather than a universal performance benchmark.

6. Forecast Generation and Performance Evaluation

After model training, the forecasting workflow generated predictions for the reserved seasonal periods.

QR ALPS provides visualization dashboards that allow users to compare forecast results with historical observations and analyze forecasting errors.

The evaluation process included:

• Actual electricity consumption
• Forecast consumption
• Point-level forecast errors
• Mean Absolute Error (MAE)
• Mean Absolute Percentage Error (MAPE)

This provides both a high-level view of model accuracy and the ability to investigate individual forecast periods.

Spring

QR Forecaster Data Visualization Dashboard
Actual, Forecast and Accuracy Hourly Data. May 18th – May 31st.

Summer

QR Forecaster Data Visualization Dashboard
Actual, Forecast and Accuracy Hourly Data. Aug 18th – Aug 31st.

Autumn

QR Forecaster Data Visualization Dashboard
Actual, Forecast and Accuracy Hourly Data. Nov. 17th – Nov. 30th.

Winter

QR Forecaster Data Visualization Dashboard
Actual, Forecast and Accuracy Hourly Data. Dec.18th – Dec 31st.

These differences demonstrate that forecasting difficulty can vary considerably by season even when the same general model architecture is used.

The results are specific to the historical Sample Town hotel dataset, forecast horizon, features, model configuration, and evaluation methodology used in this case study. They should not be interpreted as guaranteed accuracy levels for other hotels or markets.

7. Why Hotel Load Forecasting Requires Specialized Modeling

Hotel electricity demand presents a distinct forecasting challenge because its operational profile differs from many other customer segments.

Consumption may be influenced by:

• Guest activity
• Room occupancy
• Restaurant and food-service operations
• Heating and cooling requirements
• Lighting
• Laundry
• Elevators and common areas
• Day-of-week effects
• Seasonal tourism patterns
• Weather conditions
• Holidays and special events

The Sample Town analysis demonstrated that demand behavior can change substantially depending on both time of day and day type.

For this reason, a forecasting platform needs to support flexible feature engineering and model configuration rather than applying a single generic forecasting formula to every customer profile.

8. Probabilistic Forecasting and Uncertainty

A single forecast value does not fully describe the uncertainty surrounding future electricity demand.

For operational energy planning, it can be valuable to know not only the expected demand but also the range of plausible outcomes.

The QR ALPS forecasting architecture can therefore be extended to probabilistic forecasting, providing uncertainty intervals around the central forecast.

For a hotel operator or energy manager, this can support questions such as:

• What is the expected electricity demand?
• How high could demand reasonably become?
• How much uncertainty exists around the forecast?
• Which forecast periods represent greater operational risk?

This allows forecasting to become a risk-management tool rather than simply a point prediction.

10. Scenario Simulation

The same AI forecasting framework can also be extended to future scenario analysis.

For example, organizations can investigate how expected electricity demand could change under different external conditions, such as:

• Higher or lower temperatures
• Different humidity conditions
• Alternative calendar assumptions
• Changes in seasonal conditions
• Other relevant external variables

Scenario simulation provides a way to explore potential future demand conditions before they occur.

This can support planning for energy procurement, operational capacity, demand management, and other energy-related decisions.