Paper2(SVR,LSTM)
1. Sustainability in Energy Management
Article on Short-Term Load Forecasting of Microgrids using machine learning algorithms.
Focus on hybrid Support Vector Regression (SVR) and Long Short-Term Memory (LSTM).
Study conducted in a rural microgrid (MG) in Africa.
Authors and Affiliations
Arash Moradzadeh, Sahar Zakeri, Maryam Shoaran, Behnam Mohammadi-Ivatloo, Fazel Mohammadi
University of Tabriz, Iran
Duy Tan University, Vietnam
University of Windsor, Canada
Correspondence Emails: arash.moradzadeh@tabrizu.ac.ir, bmohammadi@tabrizu.ac.ir, fazel@uwindsor.ca
2. Abstract
Short-Term Load Forecasting (STLF) crucial for consumers and generators.
Proposed hybrid model (SVR-LSTM) effectively forecasts loads based on input variables (household types, commercial entities).
Achieves a high correlation coefficient of 0.9901, improving upon traditional SVR and LSTM methods.
3. Introduction
Increase in electricity consumers leading to traditional power systems troubles: reliability/stability issues.
Need for increased generation capacity coupled with environmental concerns drives focus towards Distributed Generations (DGs).
Emergence of Smart Grids for energy management using smart meters.
Microgrids (MG)
MGs as regional units independent of macrogrids; 'prosumers' can also generate energy.
Energy management in MGs can enhance consumer knowledge and facilitate bilateral energy trading.
Importance of Load Forecasting
Key for energy management, influencing both operational decisions and long-term network planning.
Load forecasting strategies categorized into various approaches, including persistence models, statistical methods, ANNs, machine learning, and hybrids.
4. Methodologies
4.1 Support Vector Regression (SVR)
Origins in machine learning, specializes in regression and forecasting.
ε-SVR focuses on maintaining an error small enough to model relationships between input and output data.
4.2 Long Short-Term Memory (LSTM)
Addresses long-term dependencies in time series data; uses memory cells to handle gradient problems.
Implements multiple gates to manage information flow effectively.
4.3 SVR-LSTM Structure
Combines periodic features as SVR inputs and time series data for effective forecasting.
Works cooperatively to predict load behavior with a high correlation coefficient.
5. Simulation Results
Dataset Description
Utilizes a freely available dataset from rural MG in Sub-Saharan Africa focusing on load profiles.
Modeling of household and commercial loads conducted hourly.
Input Variables
Characteristics such as number of households and income distributions modeled.
Accurate modeling aids in deriving load profiles under various scenarios.
Evaluation Metrics
Statistical metrics include:
Correlation Coefficient (R)
Mean Square Error (MSE)
Root Mean Square Error (RMSE)
Mean Absolute Error (MAE)
Mean Absolute Percentage Error (MAPE)
Training Phase Results
SVR (R = 0.9813), LSTM (R = 0.9852), SVR-LSTM (R = 0.9956) showcased top accuracy during training.
Results affirm the superior predictive capability of the hybrid model.
6. Test Results
Test phase results validate training.
SVR (R = 0.9770), LSTM (R = 0.9809), SVR-LSTM (R = 0.9901).
The hybrid model continues to outperform traditional models.
Observational Data
Figures exhibit hourly predictions showcasing accuracy and error assessment.
7. Discussion
Addresses challenges in energy management in MGs emphasizing accurate forecasting.
SVR-LSTM model proves superior without needing meteorological data, focusing on applied conditions.
8. Conclusion
SVR-LSTM model is an efficient machine learning application for short-term load forecasting in MGs.
The model excels in generating accurate forecasts compared to other traditional methods, thus supporting effective load management.
Limitations and Research Gaps in Short-Term Load Forecasting of Microgrids
Limitations of the Current Study
Dependency on Input Variables: While the SVR-LSTM model achieved a high correlation coefficient, its performance is heavily reliant on the accuracy and variety of input variables such as household types and income distributions.
Context Specificity: The model was developed and validated using data from a rural microgrid in Sub-Saharan Africa, which may limit its applicability to different geographical and socio-economic contexts.
Lack of Consideration for Dynamic Factors: The model does not account for dynamic factors such as weather conditions or sudden changes in consumer behavior, which can significantly influence load forecasting.
Research Gaps
Integration of Meteorological Data: Future studies could explore the integration of meteorological data to enhance forecasting accuracy, particularly in identifying load changes due to weather events.
Comparison with Emerging Models: There is a need for comparative studies against other emerging machine learning models and hybrid approaches to establish benchmarks and validate the SVR-LSTM model's advantages.
Scalability and Generalization: Research should focus on scalability and the generalization of the SVR-LSTM model across diverse microgrid systems and rural-urban settings to evaluate its practicality in broader applications.
Real-Time Application: Investigating the real-time implementation of load forecasting models to adapt to changing patterns would fill existing research gaps, ensuring models are responsive and practical.