Debottlenecking cogeneration systems under process variations: Multi-dimensional bottleneck tree analysis with neural network ensemble

Highlights: • This paper extends bottleneck tree analysis for multi-dimensional problems. • An ensemble of neural networks was used for data-driven process modelling. • The technology was deployed in a multi-train cogeneration plant in Malaysia. • The method improved 54.2% in carbon emission, 46.3% in OpEx, and 59% in energy production. • The shortest average payback period was found to be 93.9 weeks. Due to lucrative economics and energy policies, cogeneration systems have blossomed in many existing industries and became their backbone technology for energy generation. With ever-increasing energy demands, the required capacity of cogeneration gradually grows yearly. This situation unveils a crawling problem in the background where many existing cogeneration systems require more energy output than their allocated design capacity. To debottleneck cogeneration systems, this work extends the bottleneck tree analysis (BOTA) towards multi-dimensional problems with novel consideration of data-driven uncertainty modelling and multi-criteria planning approaches. First, cogeneration systems were modelled using an ensemble neural network with mass and energy balance to quantify the system uncertainty while assessing energy, environment, and economic indicators in the system. These indicators are then evaluated using a multi-criteria decision making (MCDM) method to perform bottleneck tree analysis (BOTA), which identifies optimal pathways to plan for debottlenecking projects in a multi-train cogeneration plant case study. With zero initial investment and only reinvestments with profits, the method achieved 54.2 % improvement in carbon emission per unit power production, 46.3 % improvement in operating expenditure, 59.0 % improvement in heat energy production, and 58.9 % improvement in power production with a shortest average payback period of 93.9 weeks.