(59l) Modelling of Non-Conventional Streams in the Context of Circular Economy-the Case of Hydrothermal Liquefaction

The primary aim of this paper is to introduce a method of non-conventional streams modelling. Modelling is accomplished through multi-objective optimization with 4 different algorithms and its implementation in the biochemical process Hydrothermal Liquefaction (HTL) of Sewage Sludge is being studied. There are plenty of categories for optimization techniques. Two of the biggest and most common are deterministic and stochastic methods. Deterministic algorithms will always produce the same output, given a particular input of data, while on the opposite side, stochastic algorithms capitalize on random variables and results. The optimization algorithms used in this project include Firefly Algorithms (FA), Particle Swarm Optimization (PSO), Sequential Quadratic Programming (SQP), and Least Squares Optimization (LS). The two first (FA, PSO) are stochastic methods, while the other two (SQP, LS) are deterministic. Moreover, FA and PSO are evolutionary algorithms based in swarm intelligence, while SQP and LS are algebraic methods. The implementation is performed in Matlab whilst the data is inputted in an Excel Sheet and then imported into Matlab. As for the methodology of this model, first necessary data is collected and organized and then the chosen algorithm will use this input in order to solve the problem. The objective function is expressed as sum of relative differences squared and can easily be shaped according to each problem needs. The work is illustrated in the case of Hydrothermal Liquefaction is a method of liquid biomass processing. Its main product is an oily mixture called biocrude, which in fact contains a high percentage of heteroatoms, limiting it from being used in many applications. Through various methods (upgrading), it can eventually be used as a liquid fuel. For this particular process, the goal was to predict the quantitative and qualitative composition of the 3 main streams (Biomass, Biocrude, Updated Biocrude). The parameters of this problem include the elemental and the biochemical composition, density, and the Boiling Point Distribution with the only restriction being that the mass fraction of each compound lies between 0 and 100%, and the sum of them fulfils a particular target set per stream. After running each algorithm for a number of times to establish their repeatability, only FA and PSO seemed to diverge on some iterations which can be justified by the fact that FA and PSO are stochastic algorithms.

Results showed that the most efficient algorithm was SQP, as it produced optimal results, extremely fast (6-7 times faster compared to the other algorithms). Capitalizing on the form of the problem (square error regression) and the algorithm’s simplistic design, it managed to give the overall best convergence. In most of the cases, parameters 1,2,3 were easy to accurately predict, producing a small error (max 15%). However, all of the algorithms were observed to find it difficult to attain parameter 4, which can be reasonable due to the fact that it consists of many sub-objectives. It should also be noted that FA in 3 occasions failed to provide accurate results. That most probably has to do with the fact that FA is a problem-specific algorithm, thus it needs some modification before being applied to a problem. In this case, the optimal configuration for FA could not be achieved. Moreover, PSO belong to the same category as FA, but managed to solve the problem significantly better. PSO is a more general approach that can be applied with less effort to any problem, requiring little to no modification. Lastly, LS manage to have the same convergence as SQP, demanding nevertheless a lot more time.

One other thing that should be noted, is that especially for the Updated Biocrude stream, literature data is very restricted. So, at first 14 compounds was chosen to represent Updated Biocrude, producing mediocre results. Then the database was increased to 30 compounds and the results was much better.

Thus, an increase in the efficiency of the model when increasing the number of compounds can be noted. This is to say that when modelling a stream, a sufficient amount of data should be provided, in order for the model to optimally converge. However, an upper limit also exists, which in this case is predicted to be around 80-100 compounds, depending on the needs of each stream.

Results also showed that if the updated biocrude was to be fractional distilled, it would produce 24% Kerosene and 40% Gas Oil. On these grounds, considering a typical biomass flow of 100L/h and an average biocrude yield 35%, we can expect 14 L/h Gas Oil and 8.4 L/h Kerosene to be produced after proper processing.