Faith-5/smart-meter

This project evaluates five low-complexity machine learning models to predict household electricity consumption using locally collected smart meter data from a locality, Nigeria. # ⚡ SmartMeter: Electricity Consumption Prediction ## 📖 Project Overview This project evaluates five low-complexity machine learning models to predict household electricity consumption using locally collected smart meter data from a locality, Nigeria. It specifically addresses the unique challenges of forecasting in environments with grid instability (load shedding and blackouts) and identifies the most suitable model for resource-constrained mobile deployment. ## ✨ Key Features of the Web App * 📊 **Interactive Data Explorer:** Visualizes 15,000 hourly smart meter readings (Jan 2024 – Sep 2025), revealing trimodal daily patterns and the impact of grid outages. * 🧠 **Model Comparison:** Head-to-head evaluation (MAE, RMSE, R², Storage Size) of Moving Average, Linear Regression, ARIMA, Random Forest, and LSTM. * 📱 **Mobile Deployment Analysis:** Scores each model based on accuracy vs. deployability on mid-range Android/iOS devices. * 🔮 **Live Predictor:** A playground to dynamically test the winning model (Linear Regression) with custom hypothetical inputs to forecast future consumption. * 🤖 **AI Assistant:** An integrated chatbot powered by Groq (Llama 3) to answer academic and project-related questions on the fly. ## 🏆 Core Findings (TL;DR) 1. **Best for Mobile:** **Linear Regression** offers the best balance of accuracy (R² = 0.77) and compactness (Tiny 0.9 KB footprint), making it ideal for mobile deployment without runtime dependencies. 2. **Best Raw Accuracy:** **Random Forest** yielded the best accuracy (R² = 0.90) but at 48 MB, it is too large for direct mobile integration. 3. **Key Predictors:** Historical lags—specifically the same hour yesterday (Lag_24h) and the same hour last week (Lag_168h)—are the strongest indicators of future consumption.