B.

Tech Design Project Report

Project Title:
Electricity Price Forecasting and Load Optimization using Hybrid
Machine Learning Models

Submitted in partial fulfillment

of award of
[Link] Electrical Engineering
By
Priyanshu Rathour
12214093

Under the mentorship of

Dr. Ashwani Kumar

Department of Electrical Engineering

National Institute of Technology, Kurukshetra
Haryana-136119, India
August-November 2025
Acknowledgment

I would like to express my sincere gratitude to my supervisor, Dr. Ashwani Kumar, for his continuous guidance,
encouragement, and support throughout the completion of this design project. His insights and expertise in the
field of power systems, energy management and smart grid were invaluable in shaping the direction and quality
of my work.
I am deeply thankful for his mentorship, which helped me understand the practical aspects of forecasting, opti-
mization, and smart energy management. This project would not have been possible without his consistent support
and constructive feedback.
I also extend my appreciation to the Electrical Engineering Department, National Institute of Technology, Kuruk-
shetra, for providing the academic environment and necessary resources that enabled me to carry out this work
effectively.

Date: November 21, 2025 Priyanshu Rathour

Institute Roll Number:

Place: Kurukshetra 12214093

ii
Declaration

I hereby declare that the work presented in this design project report titled “Electricity Price Forecasting and
Load Optimization using Hybrid Machine Learning Models”, submitted in partial fulfillment of the require-
ments for the award of the degree of [Link] in Electrical Engineering, is my original work carried out under the
mentorship of Dr. Ashwani Kumar, National Institute of Technology, Kurukshetra.

The work presented in this report has not been submitted by me for the award of any other degree of any other
institute. I have taken care in all respects to honor the intellectual property right, and have acknowledged the
contribution(s) of others for using them for this academic purpose. I will be liable for full or partial violation of
copyright or intellectual property right, if found at any stage even after award of my degree, not my mentor(s)/
Head of Department/ Institute.

Date: November 21, 2025 Priyanshu Rathour

Institute Roll number:
12214093

I have mentored and guided the candidate throughout the Design Project period, and I certify that the report
accurately reflects the candidate’s contributions and achievements to the best of our knowledge.

Dr. Ashwani Kumar

Professor, Electrical Engineering Department
NIT, Kurukshetra

iii
Abstract

This project presents a hybrid machine-learning framework for short-, medium-, and long-term electricity price
forecasting, combined with load optimization under Time-of-Use tariff conditions. Reliable forecasting of elec-
tricity demand and pricing is essential for modern power systems, enabling better planning, cost reduction, and
demand-side management. In this work, historical household power consumption data is preprocessed and mod-
elled using a set of complementary time-series techniques, including Auto-ARIMA, Prophet, and gradient-boosted
regression. These models are used to generate forecasts across three horizons—7 days, 1 month, and 1 year cap-
turing both short-term variations and long-term seasonal trends.
The predicted load profiles are mapped to TOU tariff categories to estimate dynamic electricity costs. An opti-
mization strategy is then applied, reducing peak demand by 10 percent and shifting load into lower-tariff periods
without altering the total energy consumption. Performance is evaluated through graphical comparison of baseline
and optimized schedules, cost distribution across tariff bands, and model accuracy metrics.
Simulation results show that the hybrid approach improves forecasting reliability across all time horizons, smooths
noisy demand signals, and delivers measurable reductions in overall energy cost. The project demonstrates the
practical relevance of data-driven forecasting and optimization in smart grid environments, industrial energy plan-
ning, and intelligent home energy management systems.

iv
Contents

1 INTRODUCTION TO ELECTRICITY PRICE FORECASTING 1

2 OVERVIEW OF LOAD FORECASTING AND TOU TARIFFS 3

2.1 Fundamentals of Load Demand Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Time-of-Use (TOU) Tariff Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.3 Relationship Between Load and Price . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.4 Factors Affecting Electricity Prices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3 MACHINE LEARNING MODELS FOR ELECTRICITY PRICE FORECASTING 6

3.1 Traditional Forecasting Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.2 Auto-ARIMA Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.3 Prophet Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.4 XGBoost Regression Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.5 Comparison of Model Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.6 Why Hybrid Models? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

4 DATASET AND PREPROCESSING 10

4.1 Dataset Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
4.2 Data Cleaning and Missing-Value Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
4.3 Time-Series Indexing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
4.4 Feature Engineering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4.5 Data Visualization and Trend Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4.6 Time Series Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

5 FORECASTING METHODOLOGY 15
5.1 Short-Term (7-Day) Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
5.2 Medium-Term (1-Month) Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
5.3 Long-Term (1-Year) Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

6 LOAD OPTIMIZATION UNDER TOU PRICING 17

6.1 TOU-Based Price Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
6.2 Peak Load Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
6.3 Optimization Strategy (10% Peak Reduction) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
6.4 Baseline vs Optimized Power Profile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
6.5 Cost Reduction Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

7 SIMULATION AND RESULTS 19

7.1 Forecast Plots (7-Day, 1-Month, 1-Year) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
7.2 TOU Price Mapping Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
7.3 Cost Comparison Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
7.4 Peak Load Shifting Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
7.5 Summary of Observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

8 Applications of Forecasting and Optimization 24

v
vi CONTENTS

8.1 Industrial Energy Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

8.2 Smart Home Energy Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
8.3 Renewable Energy Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
8.4 Utility Demand Response Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
8.5 Smart Grid and AI Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

9 Future Scope 25
9.1 Integration with Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
9.2 Including Weather and Market Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
9.3 Real-Time Optimization with IoT Meters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
9.4 AI-Driven Demand Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

Final Remarks 26

Bibliography 27
Chapter 1

INTRODUCTION TO ELECTRICITY PRICE

FORECASTING

Electricity price forecasting has become a critical discipline in modern power systems, driven by the rapid growth
of smart grids, renewable energy integration, and data-driven energy management strategies. As electricity de-
mand fluctuates throughout the day due to consumer behaviour, weather changes, industrial operations, and sea-
sonal effects, forecasting models enable engineers to estimate future load patterns and corresponding electricity
prices with accuracy and reliability.

In any power system, the balance between electricity generation and consumption must be maintained in real time.
Even slight mismatches can cause voltage instability, frequency deviations, or system stress. Since electricity can-
not be economically stored at large scales, forecasting becomes the primary tool for planning generation schedules,
market bidding, transmission line loading, and consumer-side energy optimization. As a result, forecasting has
evolved into one of the most important analytical tools in electrical engineering and energy economics.

Electricity price forecasting is closely linked with load forecasting. When the demand on the grid rises—typically
during evening hours—electricity prices increase due to the activation of costly peaking generators and increased
network stress. Conversely, during off-peak hours when demand falls, prices drop significantly. This behaviour
forms the basis of Time-of-Use (TOU) pricing, where consumers are charged different tariffs depending on the
hour of the day. Forecasting these variations allows industries and households to schedule appliances, shift loads,
and reduce electricity bills without compromising comfort or production efficiency.

A flexible and data-driven approach is essential because electricity consumption is influenced by several uncertain
and dynamic factors, including temperature, humidity, day type, holidays, and renewable generation variability.
Machine learning models such as Auto-ARIMA, Prophet, and gradient-boosted regression have become widely
used due to their ability to capture nonlinear trends, seasonality, and long-term dependencies in consumption
data. These models process historical load patterns and intelligently predict future values using mathematical
relationships learned from the data.

One of the defining strengths of machine learning–based forecasting is its adaptability. Unlike traditional statistical
models that rely heavily on fixed linear assumptions, machine learning methods can adjust to changing patterns,
noise, and irregular behaviours found in real-world energy data. For example, sudden spikes in consumption
caused by extreme weather conditions or local events can be better modeled and predicted using hybrid approaches
that combine both classical time-series theory and nonlinear machine learning techniques.

Forecasting electricity prices is particularly important in regions using smart meters and TOU billing. Under TOU
tariffs, the day is divided into three major zones:

• Off-Peak Hours: Lowest tariff, typically during night and early morning.

• Shoulder Hours: Moderate tariff, representing balanced load.

• Peak Hours: Highest tariff, when electricity demand reaches its maximum.

1
2 CHAPTER 1. INTRODUCTION TO ELECTRICITY PRICE FORECASTING

Predicting the amount of energy that will be consumed in each of these periods allows engineers to estimate
electricity costs and design optimization strategies. For instance, if a model predicts a high evening peak, users
can reduce consumption during those hours and shift non-essential loads—such as electric water heaters, EV
charging, or motor-driven processes—to off-peak periods. This not only reduces the consumer’s energy bill but
also minimizes stress on transformers, feeders, and the overall power grid.

Electricity price forecasting has wide-ranging applications in energy management and smart grid technology:

• Industrial Load Scheduling: Planning high-power operations to minimize electricity cost.

• Smart Homes: Automating appliances based on predicted low-tariff hours.

• Renewable Energy Integration: Scheduling storage and power dispatch using predicted price trends.

• Utility Operations: Predicting upcoming peaks to maintain grid stability.

• Energy Trading: Supporting bidding strategies in electricity markets.

• Demand Response Programs: Encouraging consumers to shift their load during expensive periods.

In summary, electricity price forecasting forms the backbone of modern energy planning and optimization. It
combines statistical analysis, machine learning algorithms, and real-time data processing to generate accurate
predictions across different time horizons—short-term (7 days), medium-term (1 month), and long-term (1 year).
With the growth of smart grids and intelligent metering infrastructure, forecasting continues to play an essential
role in improving energy efficiency, reducing operational costs, and enabling more sustainable and reliable power
systems.
Chapter 2

OVERVIEW OF LOAD FORECASTING AND

TOU TARIFFS

Electricity consumption is highly dynamic and varies significantly with human activities, industrial processes,
climate conditions, and seasonal factors. Understanding these variations is essential for forecasting energy demand
and predicting electricity prices, especially under modern tariff systems that change according to the time of day.
This chapter provides a foundational understanding of load forecasting, the behaviour of demand profiles, and the
importance of Time-of-Use (TOU) pricing in electricity cost estimation and optimization.

Load forecasting is a critical analytical task for power utilities, industries, and consumers alike. Accurate forecasts
help prevent grid overloading, reduce operational costs, schedule generation units efficiently, and promote optimal
usage of distributed energy resources. At the consumer side, forecasting helps in strategic decisions such as
shifting electrical loads to cheaper time intervals, planning production schedules, and reducing overall electricity
expenditure.

Time-of-Use tariffs divide the day into different pricing categories based on expected demand levels. Since load
and price are directly correlated, TOU pricing is widely adopted in smart grid frameworks and aims to encourage
consumers to avoid peak hours and relieve stress on the grid. In this project, TOU pricing forms the foundation
for translating consumption forecasts into cost estimates.

2.1 Fundamentals of Load Demand Patterns

Electricity demand typically follows predictable patterns driven by human behavioural cycles and environmental
factors. A typical daily load curve exhibits the following characteristics:
• Morning Rise: Demand increases as residential and commercial activities begin.

• Midday Plateau or Dip: Domestic load decreases while industrial operations remain steady.

• Evening Peak: The highest demand occurs due to lighting, cooling/heating, and industrial load overlap.

• Night Drop: Demand decreases substantially as most activities slow down.

Seasonal variations also play a major role. Cooling demand dominates in summer, while heating demand increases
in winter. Industrial output, festival seasons, and unpredictable events can influence daily load profiles as well.
These factors must be accounted for in forecasting models to improve prediction accuracy.

Several external variables significantly affect electricity demand:

• Weather conditions (temperature, humidity, cloud cover)

• Population behaviour and working-hour patterns

3
4 CHAPTER 2. OVERVIEW OF LOAD FORECASTING AND TOU TARIFFS

• Renewable energy production variability

• Day type (weekday, weekend, holiday)
• Industrial production cycles

2.2 Time-of-Use (TOU) Tariff Structure

Time-of-Use tariffs categorize a day into multiple price intervals, each reflecting the expected load on the grid.
TOU pricing encourages demand-side load shifting and helps utilities manage the power system more efficiently.

A typical TOU structure used in this project is as follows:

• Off-Peak Hours (10 PM – 6 AM): Lowest tariff due to low demand.
• Shoulder Hours (6 AM – 5 PM): Moderate tariff as demand gradually increases.
• Peak Hours (5 PM – 10 PM): Highest tariff when demand reaches its maximum.

The main objectives of TOU pricing are:

• Encouraging energy use during low-demand periods.
• Reducing peak load stress on transformers and distribution feeders.
• Decreasing the need for costly peaking power plants.
• Supporting smarter energy usage and consumer participation.

By mapping forecasted load values to TOU tariff categories, electricity cost can be accurately estimated for dif-
ferent time horizons such as 7 days, 1 month, or 1 year.

2.3 Relationship Between Load and Price

There is a strong and direct relationship between electricity load and price. As the load increases, the cost of
supplying energy also rises due to:
• Deployment of expensive peak-generation units.
• Increased transmission losses.
• Greater system congestion and thermal stress.
• Higher fuel usage during peak demand.

Conversely, during low-load hours, electricity is cheaper because:

• Base-load plants operate efficiently.
• Transmission losses are minimal.
• Renewable energy contributions are often higher.

By forecasting demand patterns accurately, utilities and consumers can estimate future electricity prices and plan
operations to minimize expenditure. This forms the basis for the optimization strategy used in this project.
2.4. FACTORS AFFECTING ELECTRICITY PRICES 5

2.4 Factors Affecting Electricity Prices

Electricity prices depend on a variety of technical, economic, and environmental factors:

• Demand and Load Variations: The primary influence on tariff changes.

• Fuel Costs: Prices of coal, natural gas, and oil affect generation cost.

• Availability of Renewable Energy: Solar and wind generation reduce prices during high availability.

• Market Prices: Dynamics of energy exchanges such as IEX.

• Grid Stability Constraints: Congestion and system stress raise prices.

• Government Regulations: Policies, subsidies, and taxes affect tariff structures.

These factors make electricity price forecasting a complex yet essential task for both utilities and consumers.

2.5 Summary
This chapter established the theoretical foundation for understanding electricity demand patterns and TOU tariff
mechanisms. It highlighted the strong interdependence between load behaviour and electricity pricing, demon-
strating why accurate forecasting is essential for cost optimization. The next chapter discusses various machine
learning models used for short-term, medium-term, and long-term electricity price forecasting.
Chapter 3

MACHINE LEARNING MODELS FOR

ELECTRICITY PRICE FORECASTING

Accurate electricity price forecasting relies heavily on mathematical and computational models that can capture
trends, seasonality, and nonlinear behaviour in consumption data. Modern forecasting approaches use both statis-
tical time-series models and machine learning techniques to analyze historical demand patterns and predict future
values. This chapter provides an overview of the forecasting models used in this project, namely Auto-ARIMA,
Prophet, and XGBoost regression. Each model contributes unique strengths, and together they form a hybrid
forecasting approach capable of short-term, medium-term, and long-term prediction.

Machine learning models have gained prominence in energy forecasting because they can adapt to rapidly changing
consumption patterns, incorporate multiple features, and generalize well to unseen data. These models operate
by learning mathematical relationships from historical data and applying them to future time intervals. Unlike
traditional load forecasting methods, machine learning models can handle noise, missing data, and irregularities
more effectively, making them highly suitable for real-world smart grid applications.

In this chapter, we discuss the theory and operational principles of the three forecasting models used in the project,
along with their suitability for different forecasting horizons (7 days, 1 month, and 1 year).

3.1 Traditional Forecasting Approaches

Before the era of machine learning, forecasting relied primarily on statistical models such as:

• Moving averages

• Exponential smoothing

• Holt–Winters trend and seasonality models

• Autoregressive models (AR)

• Moving average models (MA)

• Combined ARMA and ARIMA models

While effective for some applications, these methods struggle with nonlinear patterns, sudden spikes in demand,
and long-term seasonality. Machine learning models address these limitations by using richer feature sets and
nonlinear modelling capabilities.

6
3.2. AUTO-ARIMA MODEL 7

3.2 Auto-ARIMA Model

The Auto-ARIMA model is an advanced statistical method that automatically identifies the best parameters for the
ARIMA (Autoregressive Integrated Moving Average) model. ARIMA works by combining:

• Autoregression (AR): Relationship between a variable and its past values.

• Integration (I): Differencing of data to achieve stationarity.

• Moving Average (MA): Dependence on past error terms.

Auto-ARIMA automates:

• Model selection,

• Parameter optimization,

• Stationarity testing,

• Seasonal component identification.

Strengths of Auto-ARIMA:

• Highly effective for short-term forecasting.

• Models seasonality and trends well.

• Performs well when data has periodic patterns.

Limitations:

• Struggles with long-term nonlinear patterns.

• Requires stationary data.

• Sensitive to sudden demand spikes.

In this project, Auto-ARIMA is used for the 7-day short-term forecast due to its stability and accuracy in modelling
short-duration temporal dependencies.

3.3 Prophet Model

Prophet is a forecasting model developed by Facebook (Meta) designed specifically for handling:

• Multiple seasonal patterns,

• Holiday effects,

• Long-term trends,

• Irregular or missing data.

Prophet decomposes the time series into three major components:

• Trend: Long-term increase or decrease in demand.

• Seasonality: Repeating weekly, monthly, or yearly patterns.

8 CHAPTER 3. MACHINE LEARNING MODELS FOR ELECTRICITY PRICE FORECASTING

• Holiday effects: Special events or temperature-driven anomalies.

Advantages of Prophet:

• Extremely robust to missing data.

• Suitable for long-term (monthly and yearly) forecasts.

• Captures nonlinear seasonality effectively.

• Automatically detects trend shifts.

Limitations:

• Short-term accuracy is sometimes lower than ARIMA.

• Does not directly model complex nonlinear interactions.

In this project, Prophet plays a crucial role in 1-month and 1-year forecasting due to its strong capability to capture
seasonal and long-term trends.

3.4 XGBoost Regression Model

XGBoost (Extreme Gradient Boosting) is a state-of-the-art machine learning model widely used for structured
data. It works by building an ensemble of decision trees, each correcting the error of the previous one.
XGBoost uses:

• Gradient descent optimization,

• Regularization for preventing overfitting,

• Tree-based decision rules,

• Parallel computation for high performance.

XGBoost is capable of modeling:

• Nonlinear relationships,

• Interactions between features,

• Sudden spikes in load,

• Long-term patterns with engineered features.

Strengths of XGBoost:

• Excellent for medium-term forecasting.

• Handles noise and outliers more effectively.

• Works well with engineered features such as lag values, rolling averages, and time-of-day indicators.

Limitations:

• Requires careful tuning of hyperparameters.

• Performs best with large datasets.

3.5. COMPARISON OF MODEL CHARACTERISTICS 9

3.5 Comparison of Model Characteristics

• Auto-ARIMA: Best for short-term forecasting with stable patterns.

• Prophet: Best for long-term trend and seasonality modeling.

• XGBoost: Best for medium-term forecasting and nonlinear data.

Using all three models creates a hybrid forecasting framework that takes advantage of each model’s strengths.

3.6 Why Hybrid Models?

A single model is rarely sufficient for all forecast horizons. Electricity demand behaves differently at different time
scales:

• Short-term data shows strong temporal dependence.

• Medium-term data includes a mix of nonlinear patterns.

• Long-term data is dominated by seasonality and trends.

A hybrid model:

• Improves accuracy,

• Captures both linear and nonlinear behaviour,

• Enhances reliability across multiple forecasting horizons.

Thus, the combination of Auto-ARIMA, Prophet, and XGBoost forms a comprehensive forecasting system suitable
for the 7-day, 1-month, and 1-year predictions required in this project.
Chapter 4

DATASET AND PREPROCESSING

Forecasting models require a clean and well-structured dataset to ensure high accuracy and meaningful results.
This chapter explains the dataset used, the preprocessing operations performed, and the feature engineering steps
applied before forecasting. Screenshots of the main preprocessing and visualization steps are included for clarity.

4.1 Dataset Description

The dataset used in this project is the Household Power Consumption Dataset, which contains minute-level
measurements recorded over multiple years. It includes parameters such as Global Active Power, Voltage, Reactive
Power, and Sub-metering values for different circuits. The raw dataset contains two separate fields for Date and
Time, which were combined into a single timestamp for time-series analysis.

4.2 Data Cleaning and Missing-Value Handling

Real-world datasets often contain missing or invalid values, and this dataset was no exception. The following
operations were performed during preprocessing:

• Detection and conversion of missing values.

• Numeric conversion of all measurement columns.

• Removal of corrupted or incomplete rows.

• Verification of duplicates.

Figure 4.1 shows the data preprocessing step where the timestamp is set as the index, numeric conversion is applied,
missing values are checked, and the final cleaned dataset shape is displayed.

4.3 Time-Series Indexing

To ensure correct chronological alignment, the combined timestamp column was used as the dataframe index.
Time indexing supports resampling, rolling-window operations, and improves model compatibility for forecasting.

10
4.4. FEATURE ENGINEERING 11

Figure 4.1: Data preprocessing: Setting datetime index, converting numeric columns, checking missing values,
and final dataset size.

4.4 Feature Engineering

To reduce high-frequency noise present in the minute-level data, the dataset was resampled to an hourly average.
This improved the stability of the models and highlighted true consumption patterns.
Additional time-based features such as hour, weekday, and weekend indicators were also extracted.

Figure 4.2 shows the resampling and visualization code, which prepares the dataset for hourly forecasting.

4.5 Data Visualization and Trend Analysis

Visualizing the resampled time series revealed clear load patterns, including evening peaks, seasonal variation,
and long-term behavioural trends. These insights confirm that machine learning models can effectively learn and
forecast future consumption.

Figure 4.3 shows the hourly Global Active Power consumption trend across several years.

This completes the preprocessing and feature engineering pipeline used to prepare the dataset for machine learn-
ing–based forecasting.
12 CHAPTER 4. DATASET AND PREPROCESSING

Figure 4.2: Feature engineering: Resampling to hourly values and visualization code.

Figure 4.3: Hourly Global Active Power Consumption after resampling.

4.6 Time Series Decomposition

To better understand the underlying structure of the electricity consumption data, the hourly series was further
resampled into daily averages and decomposed into its fundamental components. Time series decomposition helps
separate the observed signal into trend, seasonal, and residual elements, allowing clearer insight into long-term
behaviour and recurring consumption patterns.

The decomposition plot separates the daily power consumption into four key components:

• Observed: The actual measured power consumption, containing both short-term fluctuations and long-term
variations.

• Trend: The gradual upward or downward movement in electricity usage over time. For household loads,
this may be influenced by changes in occupancy, appliance usage, or seasonal habits.
4.6. TIME SERIES DECOMPOSITION 13

• Seasonal: Recurring daily or monthly patterns, such as morning/evening peaks or seasonal variations caused
by cooling and heating loads.

• Residual: The remaining component after removing trend and seasonality, representing random fluctua-
tions, unexpected spikes, or anomalies in consumption.

This decomposition provides valuable insight into how consumer behaviour, environmental factors, and seasonal
effects influence electricity usage. It helps validate whether the dataset exhibits stable patterns suitable for fore-
casting.

Figure 4.4: Python code used for daily resampling and time series decomposition.

The decomposition results confirm the presence of strong seasonality and noticeable long-term behavioural trends
in the dataset, reaffirming the suitability of machine learning and statistical methods for electricity price forecast-
ing.
14 CHAPTER 4. DATASET AND PREPROCESSING

Figure 4.5: Time Series Decomposition of Daily Global Active Power (Observed, Trend, Seasonal, Residual).
Chapter 5

FORECASTING METHODOLOGY

Forecasting electricity consumption across different time scales requires a structured and systematic approach. In
this project, three forecasting horizons were developed: short-term (7 days), medium-term (1 month), and long-
term (1 year). Each horizon captures distinct behavioural patterns and provides insights for operational, tactical,
and strategic decision-making.

5.1 Short-Term (7-Day) Forecasting

Short-term forecasting captures immediate consumption patterns influenced by:

• hourly variations,

• day-of-week behaviour,

• recent load trends,

• short-lived fluctuations in household activity.

It is most useful for daily scheduling, operational planning, and short-term cost minimization. Short-term models
emphasize recent temporal dependencies and rapid changes in demand.

5.2 Medium-Term (1-Month) Forecasting

Medium-term forecasting focuses on predicting consumption for the upcoming month. This horizon reflects:

• weekly seasonality,

• gradual shifts in usage patterns,

• climate-driven effects,

• moderate behavioural variations.

A one-month forecast aids in energy budgeting, industrial scheduling, and anticipating tariff impacts. Models used
here must balance short-term variability with recurring weekly trends.

5.3 Long-Term (1-Year) Forecasting

Long-term forecasting identifies large-scale patterns across an entire year, capturing:

• seasonal cycles,

• long-term behavioural trends,

15
16 CHAPTER 5. FORECASTING METHODOLOGY

• climate-related changes (cooling/heating loads),

• evolving household consumption habits.

Yearly forecasting supports capacity planning, tariff design, infrastructure development, and strategic decisions in
smart grid systems. Long-term models must effectively learn slow-changing, stable seasonal structures.
Together, these three forecasting horizons offer a comprehensive view of electricity consumption behaviour and
serve as the foundation for cost estimation and load optimization.
Chapter 6

LOAD OPTIMIZATION UNDER TOU PRICING

Once electricity consumption has been forecasted for different time horizons, the next step is to optimize the load
with respect to the Time-of-Use (TOU) tariff structure. TOU pricing divides the day into off-peak, shoulder, and
peak periods, each with different electricity rates. By strategically shifting consumption from high-tariff hours to
low-tariff hours, significant cost savings can be achieved without reducing total energy usage.

6.1 TOU-Based Price Mapping

The TOU pricing structure used in this project consists of:

• Off-Peak Hours (10 PM – 6 AM): Lowest tariff.

• Shoulder Hours (6 AM – 5 PM): Moderate tariff.

• Peak Hours (5 PM – 10 PM): Highest tariff.

Forecasted hourly consumption values were mapped to these pricing bands to compute expected electricity costs.

6.2 Peak Load Identification

Peak load represents the maximum electrical demand during a specific period. Identifying peak hours is essential
because:

• Peak demand corresponds to the highest tariff.

• High peaks increase stress on distribution equipment.

The goal of optimization was to reduce the height of the forecasted peak by approximately 10% while keeping
total energy (kWh) constant.

6.3 Optimization Strategy (10% Peak Reduction)

A load-shifting technique was applied, where a small portion of consumption from peak hours was shifted to
nearby off-peak intervals. Key aspects include:

• Load shifting without reducing total consumption.

• Cost reduction by avoiding high-tariff periods.

• Grid relief through reduced evening peaks.

Shiftable loads include appliances such as water heaters, EV charging, and washing machines, while essential
loads remain unchanged.

17
18 CHAPTER 6. LOAD OPTIMIZATION UNDER TOU PRICING

6.4 Baseline vs Optimized Power Profile

Two profiles were generated:

• Baseline: Original forecasted consumption.

• Optimized: Adjusted curve with reduced peaks.

The optimized profile shows:

• Lower evening peaks.

• Slightly higher off-peak consumption.

• Identical total energy usage.

• Lower cost under the TOU tariff.

6.5 Cost Reduction Results

Electricity cost was computed by multiplying hourly load with corresponding TOU tariffs for both profiles. The
optimization achieved:

• Noticeable reduction in total electricity cost.

• Approximately 10% reduction in peak demand.

• Shift of consumption toward cheaper hours.

These results demonstrate the effectiveness of combining forecasting with TOU-based load optimization in smart
energy management.
Chapter 7

SIMULATION AND RESULTS

This chapter presents the outcomes of the implemented forecasting models, TOU (Time-of-Use) tariff mapping,
cost calculations, and load optimization techniques. The results include visual comparisons, performance analysis,
cost-saving estimations, and peak load reduction insights derived from the 7-day, 1-month, and 1-year simulations.

7.1 Forecast Plots (7-Day, 1-Month, 1-Year)

Short-term (7-day), medium-term (1-month), and long-term (1-year) forecasting models were evaluated using
ARIMA, Prophet, and XGBoost. Among these, XGBoost produced the most accurate predictions.

Figure 7.1: 7-Day Energy Forecast using XGBoost: Actual vs Forecasted.

7.2 TOU Price Mapping Graphs

Using TOU tariffs, hourly load forecasts were mapped to dynamic pricing to analyze the impact of peak, shoulder,
and off-peak periods. This mapping helps visualize the cost variation and identifies opportunities for load shifting.

19
20 CHAPTER 7. SIMULATION AND RESULTS

Figure 7.2: Model Performance Comparison: ARIMA, Prophet, and XGBoost.

Figure 7.3: Baseline vs Optimized Load, TOU Price Curve, and Hourly Cost (January 2010).

7.3 Cost Comparison Tables

Cost comparison tables help evaluate the financial impact of applying optimization over weekly, monthly, and
yearly periods.

7.4 Peak Load Shifting Analysis

Load optimization was applied to reduce high demand during peak tariff hours. The results show notable reduc-
tions in both peak load and operational cost.
7.5. SUMMARY OF OBSERVATIONS 21

Figure 7.4: Weekly Optimization KPIs: Energy usage, cost, TOU band distribution, and hourly cost plots.

Figure 7.5: Monthly Bill Comparison for January 2010 under TOU Tariff.

7.5 Summary of Observations

Based on the combined forecasting and optimization framework, the following conclusions were derived:

• XGBoost provided the best forecasting accuracy with significantly lower error compared to ARIMA and
Prophet.

• TOU optimization lowered the electricity bill without reducing total energy consumption.
22 CHAPTER 7. SIMULATION AND RESULTS

Figure 7.6: Year 2010 Summary: Baseline vs Optimized Monthly Cost, Peak Reduction, and Savings.

Figure 7.7: Energy and Cost Comparison by TOU Band (kWh and Cost).

Figure 7.8: Monthly Savings (%) for Year 2010 after Optimization.

• Peak load was reduced by an average of 8–12% across different periods.

• Monthly savings ranged between 5–12% depending on seasonal demand.

7.5. SUMMARY OF OBSERVATIONS 23

Figure 7.9: Monthly Peak Reduction (%) for Year 2010.

Figure 7.10: Annual Cumulative Cost: Baseline vs Optimized Load Profile.

• The cumulative annual cost of 2010 for the optimized load profile was considerably lower.

• Load shifting from peak to off-peak hours proved highly effective under TOU tariffs.

These results demonstrate that machine learning-driven load forecasting, combined with dynamic TOU optimiza-
tion, can significantly enhance energy efficiency and cost reduction for residential and commercial systems.
Chapter 8

Applications of Forecasting and Optimization

Accurate electricity forecasting combined with TOU-based optimization supports a wide range of practical ap-
plications in residential, industrial, and utility-level energy systems. These applications improve cost efficiency,
operational reliability, and overall energy sustainability.

8.1 Industrial Energy Management

Forecasting enables industries to estimate power demand accurately, avoid peak-hour penalties, and schedule
energy-intensive activities during low-tariff periods. Optimization techniques help maintain uninterrupted pro-
duction while reducing operational energy costs.

8.2 Smart Home Energy Systems

In smart homes, forecasting predicts appliance-level consumption and helps automate schedules for devices such
as air conditioners, washing machines, EV chargers, and geysers. This supports load shifting to off-peak hours,
reducing electricity bills and improving energy efficiency.

8.3 Renewable Energy Scheduling

Solar and wind energy systems benefit from forecasting by predicting future generation and demand. Optimization
ensures better utilization of storage systems, minimizes curtailment, and improves grid integration of renewable
energy sources.

8.4 Utility Demand Response Programs

Utilities use load forecasting to anticipate peak demand and activate demand response mechanisms. Consumers
can be incentivized to shift consumption away from congested hours, reducing stress on the power grid and im-
proving network stability.

8.5 Smart Grid and AI Integration

Smart grids rely heavily on forecasting to manage distributed energy resources, microgrids, and EV charging
networks. AI-driven optimization enhances real-time grid balancing, power quality management, and autonomous
energy dispatch.

24
Chapter 9

Future Scope

The forecasting and optimization framework presented in this project can be further enhanced by integrating ad-
vanced machine learning methods, real-time sensing technologies, and external influencing factors.

9.1 Integration with Deep Learning

Deep learning models such as LSTM, GRU, and temporal convolutional networks can be incorporated to capture
long-term dependencies and complex non-linear patterns, improving forecasting accuracy.

9.2 Including Weather and Market Signals

Weather conditions, humidity, temperature, and market price fluctuations significantly influence electricity usage.
Including these external parameters can create more robust and adaptive forecasting models.

9.3 Real-Time Optimization with IoT Meters

IoT-enabled smart meters can provide high-resolution consumption data in real time. This enables continuous
monitoring, dynamic scheduling, and instant optimization based on current grid and user conditions.

9.4 AI-Driven Demand Response

AI can automate demand response programs by learning user behavior, predicting load spikes, and autonomously
initiating optimal shifting strategies. This supports self-regulating smart energy systems with minimal human
intervention.

25
Final Remarks

The forecasting and optimization framework developed in this project demonstrates the practical value of combin-
ing machine learning with TOU-based energy management. By analysing historical load patterns and applying
predictive models such as ARIMA, Prophet, and XGBoost, the system effectively captures short-term fluctuations,
medium-term trends, and long-term seasonal behaviour.
The integration of TOU pricing into the optimization stage highlights the significant cost-saving potential achiev-
able through strategic load shifting. The results show measurable reductions in peak demand, improved off-peak
utilization, and clear financial benefits without compromising total energy consumption.
This work represents a step toward intelligent, data-driven energy systems capable of adapting to user behaviour,
dynamic tariffs, and evolving grid conditions. With advances in AI, IoT sensors, and smart meters, such forecasting-
based optimization approaches will continue to play a crucial role in future smart grids and sustainable energy
management.
The outcomes achieved in this project validate the approach and provide a strong foundation for further research,
including deep learning integration, real-time control, and AI-assisted demand response mechanisms. Overall, the
study reinforces the importance of forecasting and optimization as essential tools for building efficient, resilient,
and future-ready energy systems.

26
Bibliography

[1] H. L. Willis, Power Distribution Planning Reference Book, CRC Press, 2004.
[2] A. J. Wood and B. F. Wollenberg, Power Generation, Operation, and Control, 3rd ed., Wiley, 2013.
[3] S. Makridakis, S. Wheelwright, and R. Hyndman, Forecasting: Methods and Applications, Wiley, 1998.
[4] J. Taylor and D. Bunn, “A Quantile Regression Approach to Generating Prediction Intervals,” Energy Fore-
casting Journal, vol. 4, pp. 25–40, 2013.
[5] A. Hyndman and G. Athanasopoulos, Forecasting: Principles and Practice, OTexts, 2021.
[6] T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical Learning, Springer, 2009.
[7] J. Brownlee, Deep Learning for Time Series Forecasting, Machine Learning Mastery, 2018.
[8] Facebook Research, “Prophet Forecasting Model,” Available: [Link]
[Accessed: Nov. 2025].
[9] T. Chen and C. Guestrin, “XGBoost: A Scalable Tree Boosting System,” Proc. 22nd ACM SIGKDD, 2016.
[10] International Energy Agency, “Time-of-Use Pricing and Demand Response,” Technical Report, 2020.
[11] U.S. Department of Energy, “Advanced Metering Infrastructure and Smart Grids,” DOE Publications, 2021.
[12] Kaggle, “Household Power Consumption Dataset,” Available: [Link]
electric-power-consumption-data-set?resource=download. [Accessed: Nov. 2025].

27