The Challenge

Our client is a cellulose processing and fibre manufacturing company in India. They wanted to understand the performance decay of heat exchangers in their plant and accurately predict the steam consumption of flash evaporators. The performance decay of heat exchangers directly affects the overall process efficiency, leading to potential energy wastage and decreased productivity. Meanwhile, by accurately predicting steam consumption, the client can optimise resource allocation and improve overall plant operations. 

The primary objective of our client is twofold: 

• First, to gain a comprehensive understanding of the performance decay of heat exchangers involves analysing factors contributing to efficiency decay and developing models to quantify this decay.
  
  
• Second, leverage this understanding to predict the steam consumption of flash evaporators accurately. By achieving this, they could optimise energy utilisation and enhance overall plant operations.
  
  

The Strategy

1. Identifying factors affecting heat exchanger decay 

We have identified factors that could potentially impact heat exchanger decay. Our task involved creating an exhaustive list of these factors, aiming for a comprehensive understanding of elements influencing the deterioration of heat exchangers.

2. Understanding the cellulose manufacturing process

Upon comprehending the manufacturing process and identifying the effect on the efficiency of the heat exchangers at various stages, the analytics team mapped the process to the data points received from the client.

3. Implementing machine learning models

Multiple machine learning models were implemented. The best model, which predicted the steam output of the exchangers with >95% accuracy, was selected.

The Execution

1. Exploratory data analysis (EDA) and operating recommendations 

A period of 1 year was considered for the model, and missing data was identified. Downtime periods were identified using derived rules, and the data was segmented into multiple cycles. In total, 8 cycles were identified, where each cycle starts after the cleaning of the H.E and continues until the F.E is stopped for cleaning again. 

Outlier treatment, along with correlation analysis, was performed to understand the relationship between multiple variables, leading to valuable insights. Recommendations on ideal operating ranges were provided to maximise the operating cycle and increase efficiency. 

2. Implementing machine learning models

We identified efficiency decay bands by analysing multiple operational cycles. Machine learning models such as random forest, XGBoost, and Neural nets like FCNN were implemented to predict the steam efficiency of the cycles. The best model was selected based on accuracy and variance.

The Results

• Steam consumption can be predicted with over 93% accuracy for a forecast period of 7 days in advance
• Optimisation in steam consumption can be achieved by providing proactive suggestions for evaporator settings