GenAI Weather-Driven Store Analysis System: A Retail Intelligence Solution Combining Machine Learning Prediction and Generative AI Interaction

This article introduces a retail analysis system combining the LightGBM machine learning model and a generative AI interaction interface, developed specifically for Valvoline auto service stores. It predicts oil change business volume by analyzing weather data, providing intelligent decision support for store managers. The system achieves a prediction accuracy of R²=0.830 in tests on real data from 439 stores, revealing the deep correlation and dynamic patterns between weather and business volume.

Traditional operational decisions in the auto service industry rely on experience and intuition, lacking data-driven analysis. This project conducts research on Valvoline's 439 stores (covering 19 U.S. states), analyzing 579,000 store-day historical data points. It verifies the significant impact of weather factors on oil change business volume (95% statistical confidence level). The core issue is to quantify the degree and patterns of weather's impact on business volume.

The system includes three main modules: 1. Machine Learning Prediction Engine: Built 6 production-level models based on LightGBM (batch prediction, forward prediction, quantile models, etc.), with R²=0.830 in 2023 tests; 2. Generative AI Interaction Interface: OpenWebUI + Ollama local LLM, supporting natural language queries (e.g., 7-day predictions, weather impact analysis); 3. FastAPI Server: Provides 7 RESTful API endpoints (health check, model listing, chat completion, etc.), with a modular design for easy expansion.

1. Negative weather impacts: Heavy rain (-3.06%), strong winds (-2.34%), heavy snow (-2.18%), etc., all lead to a decrease in business volume; 2. Lag effect: Demand rebounds 1-3 days after heavy snow (up to +5.65%), no rebound after heavy rain, and demand shifts forward before extreme rainfall (previous day +2.57%); 3. Store sensitivity differences: Rain-sensitive stores see a drop of up to 23.9%, snow-sensitive ones up to 49.7%; 4. Customer type differences: Corporate customer business increases by 4.0% on rainy days, while individual customer business decreases by 3.8%.

Data sources include Valvoline's historical operation data (2018-2022), store GPS information, real-time/historical weather data (Open-Meteo, Meteostat API). Feature engineering covers time, lag, weather, and combined features. Model training follows a standard process, with validation in 2022 and testing in 2023, and 27 automated tests to ensure quality. Deployment uses Docker containerization, and local LLM (Ollama) ensures data privacy and low latency.

Operational decision support: Optimize staff scheduling, inventory management, marketing timing selection, and customer communication; Strategic analysis: Evaluate store performance, optimize network layout, and quantify weather risks; Technical transferability: Applicable to other offline industries affected by weather, such as auto services, retail, and logistics.

This system integrates machine learning and generative AI, transforming complex models into easy-to-use decision tools. Its core contributions are quantifying weather impacts, revealing dynamic patterns, identifying sensitivity differences, and providing a complete technical solution. Future directions include real-time data stream processing, causal inference, multi-modal data fusion, and personalized recommendation alerts.