Model Validation
From the optimal and near-optimal solutions of our various plant-scheduling scenarios, we can see that the overall prices of electricity per kilowatt hour in our model results are below the price of electricity for Missouri consumers, validating the accuracy of our model and suggesting that we made good assumptions regarding certain O&M and startups costs for the KCPL plants. Another validation comes from our model’s ability to adjust to changing fuel prices. Volatility in fuel prices is of high interest to utility providers whose profits hinge largely on such factors. The third weekly scenario showed significant use change of natural gas plants for KCPL when we lowered gas prices, highlighting the robustness of our model. Our model accomplishes its initial goal of mirroring demand to reduce overproduction and underproduction, with a max error measure between demand and output of 212 MWh amongst all of our scenarios.
Opportunities for Further Research
The primary constraints imposed on our research were a lack of both access to data and computing time. With access to both of these resources, we would be able to produce a more robust forecasting and allocation model. Unfortunately, our demand forecasting relied on a scaled version of time series data in England. Though many of the same daily and weekly trends were observed, obviously it would be better to have demand data for the region being focused on in our research. Another instance of lack of data was with knowledge of KCPL’s plants. Many of the parameters used in our model were simply based off of the size and type of KCPL’s plants, rather than direct numbers from the provider itself.
Computation time was yet another constraint on the extent of our research. Despite making significant simplifications to our model to reduce the number of constraints in our model, weekly scenarios all resulted in having to stop the solver and use a near-optimal solution. This resulted in us having to limit the scope of our research. Clearly, our forecasted demand that we used in our unit commitment problem is not a deterministic quantity, and the assignments obtained from running the model once are actually pulled from a distribution of possible results. Were more computation time available, we would have been able to obtain statistics such as mean and variance of production costs for our different scenarios, possibly leading to new insights.
Computation time was yet another constraint on the extent of our research. Despite making significant simplifications to our model to reduce the number of constraints in our model, weekly scenarios all resulted in having to stop the solver and use a near-optimal solution. This resulted in us having to limit the scope of our research. Clearly, our forecasted demand that we used in our unit commitment problem is not a deterministic quantity, and the assignments obtained from running the model once are actually pulled from a distribution of possible results. Were more computation time available, we would have been able to obtain statistics such as mean and variance of production costs for our different scenarios, possibly leading to new insights.