Problem Description

A major challenge in the pool-chemical industry is properly managing the supply chain. For vendors who are regularly supplying customers anticipating usage is key. For customers, running low on pool chemicals can lead to negative health consequences. One of the leading pool-chemical suppliers approached us about helping them to predict bleach usage in pools across the country. The setup:

• A fraction of the pools were equipped with IoT sensors, with a system in place to collect bleach levels in the dispensing tank
• For the remaining pools a delivery driver would manually record the remainig bleach upon delivery
• The bleach release system dispenses the chemical to keep it at recommended levels in the pool
• Bleach usage can be affected by the utilization of the pool
• Bleach usage can also be affected by factors such as weather and holidays

The forecasting problem was non-trival, as it consisted of two types of data. High quality and dependable sensor data and error-prone manually collected data.The amount of available data for any given pool varied.

Reliancy Solution

We implemented an end-to-end solution which gathers the needed data, generates the forecasts and the uploads back the results into an ERP system. These were the steps of the pipeline:

• Data Ingestion

  Delivery data and sensor data is loaded from an ERP system. Weather data is gathered from a government website.

• Data Pre-Processing and Feature Engineering

  Delivery data is interpolated from once every few days to daily. Sensor data is smoothened. Features for timeseries modeling are extracted. This includes temperatures and percipitation.

• Model Selection and Forecasting

  For each pool a collection of different approaches is evaluated. The best model is used to predict pool usage into the future.

• Flagging Of Low Confidence Predictions.

  The final results are uploaded into an ERP system. Predictions which have a low confidence are flagged as unusable. Those results are not to be used in planning future deliveries.

Impact

Our solution significantly improved the process of cost-estimation:

• Improved Planning Of Deliveries

  The forecasts that were obtained with a higher degree of confidence provided input data for a delivery optimization tool. It allowed the clients to accordinly adjust future deliveries.

• Actionable Insights

  The forecasts revealed some cases of odd usage data that were tied to a problem. Furthermore the quality of the predictions informed what non-sensor pools should be equiped with sensors, as the client continued to invest in IoT capabilities.

• Data Quality Issues Revealed

  On several occasions our pipeline flagged unexpected data. In most cases it ended up being improperly entered data, which was eventually addressed.