Demand forecasting for supply chain optimization and logistics planning

Overview
Our client, one of India’s leading cement manufacturers with over 20,000 dealers and sub-dealers and with a presence in 22 states, was facing challenges in inventory planning and distribution. Despite high production capacity, the client was facing supply shortages and frequent stock-outs. Additionally, due to lack of robust forecasting and inventory planning capabilities, the client’s logistics team was unable to pre-arrange distribution vehicles in advance leading to high Order Execution Time (OET).
TransOrg Analytics developed a machine learning based demand
forecasting solution to predict daily demand at a ‘plant-district’ pair level. These daily forecasts are then used by the client’s logistics team to pre-arrange
vehicles reducing the overall Order Execution Time (OET) to 4hrs.
Initially, twenty-five ‘plant-district’ combinations were shortlisted contributing ~ 75% of the total regional cement production. Eventually scaling up the solution to all the plants across 50 districts.
Solution
TransOrg used various AWS components such as AWS Glue, AWS Lambda, AWS Sage Maker and AWS QuickSight to develop and deploy the solution.
Input data in the form of MS Excel files were imported to Amazon S3 bucket and included:
- Daily plant production data at a district level
- Weather and macro-economic data
The data was then categorized, cleaned, and transformed using AWS Glue to make it fit to build machine learning models.
A fully managed machine learning service, AWS Sage Maker, was used to build, train, deploy and analyse models all in the same application. Metrics and forecast results generated by models were delivered to S3 bucket.
Further, the forecast results file in the S3 bucket is imported in AWS QuickSight and dashboards are created with the detailed analysis and calculation of KPIs such as total demand, total prediction of each district in a particular time-period and MAPE score for all the quantiles.

Impacts
Achieved 85% monthly accuracy in 40 districts
Achieved 72% monthly accuracy in all 50 districts
Reduced average order execution time to 4 hours