Streamlit-based Power Load Forecasting Web Application: Multi-Model Comparison and Hybrid Strategy

This article introduces a power load forecasting web application built with Streamlit, integrating models such as Persistence, Linear Regression, XGBoost, and Bi-LSTM, and proposes an XGBoost+Bi-LSTM hybrid strategy to achieve high-precision forecasting. The application covers the entire workflow from data upload to model evaluation, supports multi-model comparison and interpretability analysis, and targets power system operators and energy researchers.

Power load forecasting is a core issue in energy management; accurate forecasting is crucial for grid dispatching, energy trading, and system stability. Short-term forecasting (from hours to days) is particularly critical for day-ahead dispatching and real-time control, but traditional methods struggle to capture the complex nonlinear patterns in load data. This application focuses on power system operators and energy researchers, helping them quickly upload data, compare model performance, understand model strengths and weaknesses, and obtain interpretable results.

The project uses a Python technology stack:

• Web framework: Streamlit (quickly converts Python scripts into interactive applications without front-end experience);
• Data processing: pandas (cleaning, feature engineering);
• ML models: scikit-learn (Linear Regression), XGBoost (gradient boosting tree), TensorFlow/Keras (Bi-LSTM);
• Visualization: matplotlib, seaborn (generate data and result charts).

Data input supports CSV upload (including timestamp and load columns, auto/manual recognition) and synthetic dataset generation (one-year hourly data). Preprocessing includes dataset summary (number of rows/columns, missing values, statistics). The EDA module provides hourly load curves, monthly demand analysis, and day/week pattern visualization. Feature engineering extracts time features (hour, day of week, month), lag features (lag_1/24/168), rolling statistical features (mean, standard deviation), and missing values are handled by deletion.

Five models are implemented:

1. Persistence: Baseline model, where the predicted value is the actual value at the previous moment;
2. Linear Regression: Based on scikit-learn, assuming a linear relationship between features and the target;
3. XGBoost: Parameters n_estimators=200, learning_rate=0.05, max_depth=6; good at capturing nonlinear interactions;
4. Bi-LSTM: Input 24-hour sequence, bidirectional LSTM (64 units) + Dropout (0.2) + fully connected layer; Adam optimizer, MSE loss, MinMax scaling, early stopping mechanism;
5. Hybrid Model: Average the results of XGBoost and Bi-LSTM; resolve the start point difference of sequence input through timestamp alignment.

An 80/20 time series split is used (to avoid data leakage), with evaluation metrics MAE, RMSE, and MAPE. MAPE performance on the gbload1.csv dataset: Persistence (4.69%), Linear Regression (2.78%), XGBoost (1.30%), Bi-LSTM (1.34%), Hybrid Model (1.12%). Result display includes comparison tables, MAPE bar charts, prediction curves, feature importance, and error distribution.

Deployment methods:

• Local run: Clone the repository → virtual environment → install dependencies → streamlit run app.py;
• Streamlit Cloud: Connect GitHub repository → specify app.py as entry point;
• Hugging Face Spaces: Create a Streamlit Space → upload files for automatic deployment. Optimization suggestion: If deep learning models are not needed, remove the TensorFlow dependency to reduce resource usage.

This application is an excellent machine learning engineering practice, fully implementing an end-to-end workflow, integrating multiple models, and achieving optimal performance through a hybrid strategy. Its advantages include practical technology selection, comprehensive model design, rich visualization, and flexible deployment. It is an excellent reference for time series forecasting developers and a practical tool for power industry analysts.