(734c) Application of a Data-Driven Modeling Approach to a Large-Scale Power Plant

Most conventional fossil-fuel-based power plants are designed to operate in baseload operation with small variations in the power output [1]. With the increasing contribution of renewable energy sources to the power grid, the conventional power plants are experiencing more fluctuations in the grid. This is mainly because most of the renewable energy sources are intermittent in nature [2] and conventional power plants must ramp up/down to provide room for the renewable power to be utilized or to compensate for them when they become unavailable. The sharp ramps in the power output of conventional sources result in increased inefficiencies in the power plant operation, which results in increased generation of pollutants such as NOx. It is, therefore, critical to optimize the operation of the power plant when it undergoes transient circumstances.

In this research study, a large scale coal-fired power generation unit in the United States was considered to demonstrate the benefits of dynamic optimization, as opposed to the steady-state optimization approach that is currently implemented in this unit. Currently, power production from renewable sources as well as the electricity demand are forecasted a few hours ahead of time. This serves as an opportunity to plan the operation of the power plant ahead of time such that NOx generation in the plant is minimized while it undergoes fluctuations. The first step in implementing a dynamic optimization approach is to develop a dynamic model of the plant. This presentation provides the results of a dynamic model of the plant as opposed to a steady-state approach that is the currently utilized in the plant. The power plant considered in this analysis has hundreds of variables that makes it very complex to develop a first-principles model of the plant. Thus, a data-driven modeling approach was considered in this study. The method used to develop the dynamic model of the plant is based on a recurrent neural network (RNN) [3]. An RNN model uses neurons that have nonlinear activation functions, which is used to map inputs to outputs. An RNN model takes advantage of some of the past values of measurements (e.g. NOx in this study) that are fed back to the model; thus, making it dynamic in nature. Measurement feedback also serves as a recalibration of the model. An RNN model is also capable of predicting into the future of the plant for as many time steps ahead as there is forecasted data for power demand and renewable power sources.

To develop an RNN model of the plant, it should first be trained i.e. the different weighting and bias factors should be calibrated such that the model predictions for target outputs (i.e. NOx in this analysis) should have the minimum deviations with the corresponding historical data. To achieve this goal, historical data for all decision variables and target NOx values for over four months of the plant operation was used. Additionally, different numbers of neurons as well as model orders were investigated to balance the model prediction accuracy and its complexity. The order of an RNN model refers to the number of past measurements that are fed-back to the model. To compare the accuracy of the models with different neurons and orders, coefficient of correlation (R) was used as the performance metric.

The optimal number of neurons that can best represent the nonlinear system of this power plant was obtained after a sensitivity analysis. Subsequently, the training of the RNN was implemented for orders 1 to 3. Additionally, a static neural network was also trained with the same historical data to provide a comparison basis for the RNN models. In a static neural network, none of the past values of NOx is given to the model; thus, essentially making the neural network a steady-state model. Training of all models was implemented in the MATLAB programming language. The R-values associated with the static and RNN models of order 1 to 3 were observed to be 0.85962, 0.92919, 0.9295, and 0.92816, respectively. While all models show a strong relationship between the model prediction and historical data, the R-value associated with a recurrent neural network was always greater than 90 and greater than a static neural network. This demonstrates that a dynamic model can better represent the historical data than a static model. Among the different orders considered for the recurrent neural network, a marginal improvement was also seen by increasing the model order. Because increasing the model order increases the computational time to train and use the neural network model, a model order of 2 was selected for the further analysis of this study to balance the accuracy and computational complexity of running an RNN model.

The trained neural network was then used to predict NOx for future time steps. A sensitivity analysis was implemented on the length of the prediction horizon for both the static and RNN models. Consequently, it was observed that smaller prediction horizons resulted in higher and lower R- and root mean square error (RMSE) values, respectively, when compared to the historical NOx data; i.e. the model can more accurately predict the future operation of the plant with shorter horizons. This conclusion was expected because the further the model has to foresee into the future, it has to predict additional fluctuations that the actual plant has undergone through this period. A 3-hour prediction horizon was therefore selected for this analysis, which resulted in accurate prediction of NOx over the horizon while also enabling to demonstrate the benefits of dynamic simulation in comparison to a static model. In this study, historical data for the gross power output of the plant was used as a given input to the model for the desired prediction horizon and the NOx prediction of the model was compared with the historical NOx generation of the plant during the same period. While the static model does not require any measurement update for the 3-hour prediction horizon, this analysis assumed that the actual NOx data from the plant was available to the RNN model every 5 time steps (with 5-minute interval). For a 3-hour ahead prediction of NOx, the R-value for the static model was obtained to be 0.7796 while a value of 0.8298 was observed for the second order RNN model. This difference between the static and second order RNN models represents a 6.44% improvement in the accuracy of model prediction in favor of the RNN model. Improvement of the predictions with a RNN model became more obvious when a 24-hour prediction horizon was considered. With a 24-hour prediction horizon, it was observed that the R-value of the RNN model improved 26.92% over the static model. This is because the plant experiences more transient circumstances over a 24-hour horizon and the RNN model better encounters for these transient operations over this period. These results demonstrate the value of dynamic simulation in comparison to steady-state simulation, when the plant undergoes transient operations such as those encountered when renewable sources are being utilized in the power grid. The future research in this study includes developing a dynamic optimization algorithm for the transient operation of the plant as well as improving the accuracy of the second order RNN model.

References

[1]	"Dynamic optimization of a hybrid system of energy-storing cryogenic carbon capture and a baseline power generation unit," Applied Energy, pp. 66-79, 2016.
[2]	"Intermittency and Value of Renewable Energy," Journal of Political Economy, pp. 124, (4), 2016.
[3]	Principles of Artificial Neural Networks, World Scientific Publishing Company, 2013.