GridGenius Paper

Abstract

Bangladesh faces critical challenges in balancing electricity generation with demand, leading to resource wastage, high costs, and load shedding. We present GridGenius, a novel, fully deployed AI-powered platform for accurate and explainable daily electricity demand forecasting.

The system uses a hybrid model integrating classical ML (Random Forest, XGBoost) with a custom Transformer regressor, achieving R² scores up to 0.89. A unique dataset from 1,800+ daily BPDB reports (2020-2024) was created, enriched with holiday, temperature, and seasonal data.

GridGenius incorporates a RAG pipeline with a LLM for transparency, allowing natural language queries about forecasts. Deployed as a web app, it offers a scalable, transparent solution for smarter grid management.

Introduction

Bangladesh's energy sector is marked by a persistent struggle to align electricity generation with national demand. This imbalance results in frequent load shedding, inefficient resource use, and significant economic consequences. Despite generation capacity growth, the BPDB [1] still grapples with demand-supply mismatches, especially during peak periods and seasonal transitions. A key bottleneck is the lack of dynamic, interpretable, and real-time forecasting systems [3] to guide grid optimization efforts.

Traditional forecasting approaches, including statistical models (ARIMA) and basic ML techniques, often fall short. They may struggle to generalize across time scales, ignore crucial external factors like holidays or weather patterns, and lack the scalability needed for daily operational deployment. This highlights the need for an intelligent, adaptable system providing accurate daily forecasts coupled with clear, understandable explanations.

Our Contribution: GridGenius

GridGenius is designed to fill this gap by integrating novel, state-of-the-art AI techniques:

A Hybrid Modeling Strategy: Combines established ML algorithms (Random Forest, XGBoost) with a custom Transformer architecture [2].
Novel & Enriched Dataset: A comprehensive daily record (2020-2024) from BPDB reports [1], enhanced with temperature, holidays, and engineered features.
Explainable AI (XAI) Core [3]: Utilizes a RAG [4] pipeline with Llama 3.1 [14] for transparent, natural language interaction.
Full Deployment: An novel interactive web application providing forecasts, visualizations, and conversational insights.

BPDB Reports Processed

Peak R² Score Achieved

Years of Data Analyzed

Related Work & Research Gaps

Energy demand forecasting has evolved, but gaps remain. We reviewed relevant studies to position GridGenius:

Review of Existing Approaches(Click to Expand)

Classical ML: Hossain [5] used XGBoost on partial BPDB data, lacking features/explainability. Haque [7] applied LightGBM to limited apartment data. Wang [6] explored CNN-LSTM but ignored external factors. Haque [8] used KNN, facing overfitting and lack of daily detail.
Transformers: While powerful [2, 9] (e.g., Autoformer [10]), their application to operational energy forecasting with specific regional data is sparse.
Explainability (XAI): Often limited to metrics. LLM-powered RAG [4, 11] is emerging but not widely adopted in this domain.

Table I: Summary of Related Work (Key Limitations)
Paper	Approach	Limitations Noted
Hossain [5]	XGBoost	Limited features, no explainability
Haque [7]	LightGBM	Small dataset, not generalized, lacks external factors
Wang [6]	CNN-LSTM	Ignored seasonality & external features
Haque [8]	KNN	Lacked daily granularity, overfitting
Wen [9]	Transformer Survey	No real-world energy system applications
Wu [10]	Autoformer	Tested on synthetic/benchmark data only

                 Key Research Gaps Addressed by GridGenius:
                   Created comprehensive, multi-year, daily BPDBBangladesh Power Development Board dataset.
 Implemented robust feature engineering (holidays, temp, season, gap).
 Developed and tested hybrid ML/Transformer models tailored for this task.
 Integrated interactive XAIExplainable Artificial Intelligence via RAGRetrieval-Augmented Generation pipeline.
 Deployed a full-stack, user-friendly system.

             

Methodology

GridGenius follows a systematic pipeline:

Data Collection & Preprocessing
Exploratory Data Analysis (EDA)
Feature Engineering
Model Development & Training (Hybrid Approach)
Explainability (RAG-LLM Integration)
System Deployment

Figure 1: GridGenius System Architecture Overview. (Click to Enlarge)

A. Dataset Collection & Description

Unique dataset scraped from 1,800+ daily BPDB reports (Jan 2020 - Mid 2024). Includes Max Demand (MW), Max Generation (MW), Date, enriched with Daily Avg Temp (°C) & National Holiday flag.

B. Exploratory Data Analysis (EDA)

EDA revealed key patterns:

Figure 2: Temperature Distribution histogram showing a peak frequency around 30°C — Fig 2: Temp. Peak ~30°C

Figure 3: Demand Distribution histogram showing a broad spread, slightly skewed right — Fig 3: Broad Demand Spread

Figure 6: Scatter plot showing Demand (Y-axis) generally increasing with Temperature (X-axis), especially above 25°C — Fig 6: Demand Rises with Temp

Figure 7: Box plot comparing Demand on Holidays vs Non-Holidays, showing significantly lower median and range on holidays — Fig 7: Lower Demand on Holidays

C. Feature Engineering & Data Iterations

Features engineered: Demand-Gen Gap, Season Classification, Holiday Flag. Preprocessing: Outlier handling (IQR, Z-score), Scaling (MinMaxScaler, StandardScaler). Four dataset variants created for robustness testing.

D. Modeling Pipeline

Models explored: Linear Regression, Random Forest (Top Classical), XGBoost, SVR, Transformer (Custom DL). Tuned RF/XGBoost via RandomizedSearchCV.

E. Primary Model: Transformer Regressor

Custom Transformer [2] with multi-head self-attention, feed-forward layers, positional encoding, dropout. Trained with Adam [12] optimizer & MSE loss.

Table II: Transformer Implementation Details (Selected)
Parameter	Value
Optimizer	Adam [12]
Loss Function	MSE
Training Epochs	50
Key Hyperparameters Tuned	num_heads, model_dim, num_layers, dropout, learning_rate

F. Explainability via RAG-LLM

Achieved using RAG [4]: ChromaDB [13] stores context (docs, stats) embedded via Google/HF models [26, 27]. Llama 3.1 [14] on Groq [15] generates answers based on retrieved context, enabling natural language queries.

G. Deployment Architecture

Modern stack: FastAPI [16] backend on Railway [28], Responsive frontend on Vercel [17], persistent ChromaDB [13], Groq [15] for LLM.

Figure 10: Screenshot of the GridOracle Prediction Tool interface showing input fields (date, temp, holiday) and predicted demand output — Fig 10: GridOracle Tool

Figure 11: Screenshot of the Visualization Dashboard showing time series charts of demand, generation, and temperature — Fig 11: Insights Dashboard

Figure 12: Screenshot of the GridGenius Chatbot interface showing a user query and an LLM-generated explanation — Fig 12: GridGenius Chatbot

Results and Experiments

A. Experimental Setup & Tools

Hybrid setup (Colab T4 GPU, RTX 3060, M1 Mac). Libs: Sklearn [18], TF [19]/Keras [21], PyTorch [20], NumPy [22], Pandas [23], Matplotlib [24]/Seaborn [25], FastAPI [16], Chroma [13], etc.

B. Evaluation Metrics

Measured by R², MAE, MSE, RMSE.

C. Overall Model Performance

Random Forest & XGBoost excelled among classical models (R² ≈ 0.90).

Table III: Detailed Evaluation (Best Model Per Variant)
Dataset Variant	Best Model	R²	MAE	RMSE
GGDataset_a (MinMax + IQR)	Random Forest	0.898	0.046	0.065
GGDataset_b (MinMax + Z-Score)	Random Forest	0.892	0.048	0.067
GGDataset_c (StdScaler + IQR)	Random Forest	0.898	0.226	0.313
GGDataset_d (StdScaler + Z-Score)	Random Forest	0.894	0.227	0.318

Figure 13: Plot showing Random Forest predicted demand (Y-axis) vs actual demand (X-axis), points cluster closely around the diagonal line indicating good fit — Fig 13: Random Forest Perf.

Figure 14: Plot showing XGBoost predicted demand (Y-axis) vs actual demand (X-axis), similar good fit to Random Forest — Fig 14: XGBoost Perf.

D. Hyperparameter Tuning

RandomizedSearchCV significantly boosted RF/XGBoost performance.

E. Transformer Model Results

Strong generalization (R² ≈ 0.82).

R² Score

≈ 0.82

MAE

≈ 0.06*

RMSE

≈ 0.08*

Figure 15: Plot showing Transformer predicted demand (Y-axis) vs actual demand (X-axis), points show good correlation but slightly more scatter than RF/XGBoost — Fig 15: Transformer Perf.

Figure 16: Plot showing Transformer training and validation loss curves decreasing over epochs, indicating learning — Fig 16: Transformer Loss

*MAE/RMSE likely reported on scaled data.

F. Ablation Study: Scaling

MinMaxScaler slightly better for classical models; Transformer less sensitive. Z-score impact minimal vs IQR for outlier handling.

G. XAI Chatbot Results

RAG chatbot effectively answered complex queries in real-time (<1s latency via Groq), enhancing transparency.

Conclusion and Future Work

GridGenius presents a significant advancement in energy demand forecasting for Bangladesh. By integrating a novel dataset, hybrid modeling, and an interactive RAG-based XAI system, it addresses key limitations. The platform achieves high accuracy (R² ≈ 0.89) and offers unprecedented transparency.

The deployed system provides a practical, scalable tool for grid planners, offering a blueprint for intelligent energy management in developing economies.

Future Work Directions(Click to Expand)

Data Expansion: Seek pre-2020 BPDB data.
Model Refinement: Optimize Transformer (tuning, TCNs [30], Informer [31], multivariate attention [32]).
LLM Enhancement: Improve RAG (prompts, retrieval, LoRA [29]).
Feature Integration: Add economic/operational data.
Platform Evolution: Mobile-first UI, public APIs.

References

Bangladesh Power Development Board, "Bangladesh Power Development Board," [Online]. Available: https://www.bpdb.gov.bd/.
A. Vaswani, et al., "Attention Is All You Need," arXiv:1706.03762, 2017. [Online]. Available: https://arxiv.org/abs/1706.03762.
G. P. Reddy and Y. V. P. Kumar, "Explainable AI (XAI): Explained," 2023 IEEE eStream, pp. 1-6, doi: 10.1109/eStream59056.2023.10134984.
P. Lewis, et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks," arXiv:2005.11401, 2020. [Online]. Available: https://arxiv.org/abs/2005.11401.
M. Hossain, et al., "Short-Term Electricity Demand Forecasting of Dhaka City Using Machine Learning Approaches," arXiv:2406.06651, 2024.
Y. Wang, et al., "Short-term power load forecasting using SSA-CNN-LSTM method," Syst. Sci. Control Eng., vol. 12, no. 1, pp. 1-12, 2024. doi: 10.1080/21642583.2024.2343297.
H. Haque and M. A. Razzak, "Medium-Term Energy Demand Analysis Using Machine Learning," SSRN Electron. J., Aug. 2022. doi: 10.2139/ssrn.4197655.
H. Haque, et al., "Long-term Energy Demand Analysis using Machine Learning Algorithms," 2024 6th Int. Conf. SGRE, pp. 1-6. doi: 10.1109/SGRE60102.2024.10815643.
Q. Wen, et al., "Transformers in Time Series: A Survey," arXiv:2202.07125, 2022. [Online]. Available: https://arxiv.org/abs/2202.07125.
H. Wu, et al., "Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting," arXiv:2106.13008, 2021. [Online]. Available: https://arxiv.org/abs/2106.13008.
IBM, "Using the watsonx.ai Time Series Forecasting API to predict energy demand," IBM Developer, 2025. [Online]. Available: Link.
D. P. Kingma and J. Ba, "Adam: A Method for Stochastic Optimization," arXiv:1412.6980, 2014. [Online]. Available: https://arxiv.org/abs/1412.6980.
Chroma, "Chroma: The AI-native open-source embedding database," GitHub, 2023. [Online]. Available: https://github.com/chroma-core/chroma.
Meta AI, "Introducing Llama 3.1," Meta AI Blog, 2024. [Online]. Available: https://ai.meta.com/blog/meta-llama-3-1/.
Groq, "GroqCloud: Fast AI Inference," 2024. [Online]. Available: https://groq.com/.
S. Ramírez, "FastAPI," 2023. [Online]. Available: https://fastapi.tiangolo.com/.
Vercel, "Vercel Documentation," 2024. [Online]. Available: https://vercel.com/docs.
F. Pedregosa, et al., "Scikit-learn: Machine Learning in Python," JMLR, vol. 12, pp. 2825-2830, 2011.
M. Abadi, et al., "TensorFlow: Large-scale machine learning on heterogeneous systems," arXiv:1603.04467, 2016. [Online]. Available: https://arxiv.org/abs/1603.04467.
A. Paszke, et al., "PyTorch: An Imperative Style, High-Performance Deep Learning Library," NeurIPS, 2019.
F. Chollet, "Keras," GitHub, 2015. [Online]. Available: https://github.com/keras-team/keras.
C. R. Harris, et al., "Array programming with NumPy," Nature, vol. 585, pp. 357-362, 2020.
W. McKinney, "Data Structures for Statistical Computing in Python," Proc. 9th Python Sci. Conf., 2010, pp. 51-56.
J. D. Hunter, "Matplotlib: A 2D Graphics Environment," Comput. Sci. Eng., vol. 9, no. 3, pp. 90-95, 2007.
M. Waskom, "Seaborn: Statistical Data Visualization," J. Open Source Softw., vol. 6, no. 60, p. 3021, 2021.
Google, "Google Cloud AI Platform," 2024. [Online]. Available: https://cloud.google.com/ai-platform.
Hugging Face, "Hugging Face Transformers," 2024. [Online]. Available: https://huggingface.co/transformers/.
Railway, "Railway: Infrastructure for Developers," 2024. [Online]. Available: https://railway.app/.
E. Hu, et al., "LoRA: Low-Rank Adaptation of Large Language Models," arXiv:2106.09685, 2021. [Online]. Available: https://arxiv.org/abs/2106.09685.
S. Bai, et al., "An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling," arXiv:1803.01271, 2018. [Online]. Available: https://arxiv.org/abs/1803.01271.
H. Zhou, et al., "Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting," AAAI, 2021. [Online]. Available: https://arxiv.org/abs/2012.07436.
H. Wu, "Revisiting Attention for Multivariate Time Series Forecasting," arXiv:2407.13806, 2024. [Online]. Available: https://arxiv.org/abs/2407.13806.