Fingerprint of a Phosphorus Producing Submerged Arc Furnace A - The Limits of Dynamic Modelling

摘要

Within a phosphorus producing submerged arc furnace it was found that the continuous fluctuation of the furnace between a flowrate-driven state (high throughput) and a thermodynamic-driven state (low throughput) caused both techniques to have similar overall, predictive abilities and resulted in a linear, ARX-type, adaptive prediction model as the model of choice. Secondly, this prediction model was developed, tested and then shown to have a reasonable, 8-hour-ahead predictive accuracy (R2, coefficient of determination) of 30% (±6%) on future Pslag values. This inherent relationship exists because, at the precise moment of a Pslag prediction, the furnace already contains the metallurgical memory (input variables to the linear model) needed to ensure that some predictive possibilities will always exists all as a result of the long residence times in the furnace. Residence time, however, is not a directly adjustable variable but rather a fully dependent variable and a function of an array of interconnected and interactive variables. In fact, this applies to virtually all input variables, re-emphasising the importance of innate metallurgical memory. Thirdly, predictive control possibilities were explored by simulating the set-points of two fully independent variables used by operators to control the process: the ratio of fixed carbon-to-P2O5 and the ratio of silica gravel-to-pellets. This linear, predictive control model showed only a slight improvement with an 8-hour-ahead predictive accuracy of 35% (±7%). This highlights how ineffective current adjustments are in optimally steering the process and how difficult even incremental improvements in feed-forward and predictive control can be. Finally, it is shown that fundamental design-, samplingand process restrictions currently associated with the process will always limit the predictive and especially control accuracy or meaningfulness of any dynamic model. These restrictions include the size of the furnace resulting in long residence times, 8-hours sampling intervals, an extremely complex and interactive process and a 16% spatial analyses variation on the Pslag values the very value that the model is to predict. The point is made that, given the current status quo, even the perfect dynamic prediction model can not improve on an 8-hour-ahead prediction of 30% (±6%). This barrier can only be pierced with e.g. tidier and more frequent sampling regimes and other upfront capital investments, a decision that becomes a cost accounting exercise and that can only be taken by the management structure. An investment demanding ever-increasing attention is CFD software and it's potential to shed more light on the complex interactions within the furnace.