Downscaling and enhancement of mesoscale weather forecast model results by using different methods

The use of electricity in the world is increasing day by day with the development of technology, urbanization and industrialization. The damage caused by fossil fuels to nature and the decline of the fossil fuels leads to think alternative options more carefully; therefore the tendency to renewable energy usage is increased. The use of wind energy is increasing every year because it is renewable and not dependent on other countries resources. Potential of the field before the installation of the wind power plants is very important in terms of preventing estimated financial loss after installation. Before the establishment of the wind power plant, it is necessary to take advantage of the wind atlases for effective use while the region selection is made. For this reason, precise and accurate wind maps are of great importance in terms of the area where the wind power plant will be installed, and the effective use of the energy to be produced. Although current European and Turkish Wind Atlas (REPA) can respond to certain needs such as the changes in land use over the last 20-30 years, changes in weather patterns that are influenced by climate change, the increase in data measurement network and wind developments in the energy sector, It needs to be improved. These developing needs to be updated with respect to the high-resolution atmospheric models used for the atlas and the verification procedures of the observational data. For this purpose, the European Union project has been initiated with the aim of bringing a new and detailed wind atlas into the wind industry through the project of The New European Wind Atlas (NEWA), the European Union ERANET + program and the consortium formed by 9 member countries. In connection with this project Turkey Scientific and Technological Supported by Research Institution (TUBITAK) and Project 215M386, titled "Modeling of Wind Energy Resources on Turkey in the Scope of the New European Wind Atlas (NEWA) Project" carried out by Prof. Dr. Ş. Sibel Menteş, is ongoing. This thesis has been made within the scope of the project described above. Areas with high potentials were selected as the result of meso scale model simulations carried out within the scope of the project. Model output statistics was applied to Mersin-Mut, which is chosen as one of the highest potential regions. The relationship between the model simulation data at the selected grids and the observation point was analyzed by using stepwise regression as a variable screening method and the most effective meteorological parameters on the wind speed change were determined. Then, parameters calculated at grid points were used as input data for artificial neural networks. The aim is to minimize model errors by the applied artificial neural networks method. The study was conducted with the hourly data for 2015. For each grid, 45 parameters were extracted, with a total of 16 grids from the WRF model results at 60 m, 500 hPa and 700 hPa levels. Artificial neural networks (ANN) were applied to results after screening with Stepwise regression. Results were also analyzed via parameter elimination to see the effects of the parameters on performance. First, all the parameters were used to determine the number of neurons in the hidden layer. Then, one parameter is extracted at a time and the performance of the ANN which receives all the other parameters is analyzed. If the extracted parameter causes the performance to be unchanged or falls, it is discarded and parameters with a positive effect on performance are finally determined. In addition, 3 levels have been tested and the best result is seen when using outputs from ground level. Error analysis was performed by comparing the results of the data obtained for the test with the data of the observations. In error calculations, mean absolute error (MAE), root mean square error (RMSE) and normalized root mean square error (nRMSE) were used. One of the most widely used numerical weather prediction model, The Weather Research and Forecasting (WRF) model is used to create the European Wind Atlas by using the ERA-Interim reanalysis data set. The WRF model is a next generation mesoscale weather forecasting system that can be used for comprehensive meteorological applications. The model has mainly two model structures for research and operational usage purposes. The WRF model can be used in both hydrostatic and non-hydrostatic mode, it uses terrain following hydrostatic pressure coordinate system, and its grid structure is based on Arakawa-C grid scheme. Using the version 3.6.1 of WRF, it is running at 27 km for all of Europe and 9 and 3 km for four regions that have different climatic characteristics where sensitivity analyzes are to be carried out. WRF was run daily and weekly using two different boundary layer schemes, YSU (Yonsei University Scheme) and MYNN (Mellor-Yamada-Nakanishi-Nino). Because the MYNN scheme has the lowest error values of the results obtained daily, WRF results of the daily running MYNN was used in this study. Model output statistics (MOS) is a weather forecasting technique that allows the statistical relationship between dependent variables and independent variables obtained from the numerical models. In this study, downscaling analysis was performed for Mut and MOS were applied to the results obtained from WRF model with using Regression and Artificial Neural Networks methods. A stepwise regression analysis was performed by using the model simulation data at the selected grid points and the observation point. Most effective meteorological parameters on the wind speed change were determined by screening. Then chosen parameters were used as input data for artificial neural networks and multiple regression analysis. Artificial Neural Networks, are mathematical/computational model that try to simulate the structures and/or functions of biological neural networks. These structures, made up of artificial neuron layers connected between themselves and process the information in this network structure. Generally, ANNs are systems that adaptively change their internal weight and polarization terms according to the information flowing through them during the education process. ANNs are non-linear data modeling tools. They can learn complex relationships between input and output data or find patterns in data. High noise tolerances and non-linear structures provide great advantages. A three-layered feedforward artificial neural network trained by the Levenberg-Marquardt algorithm is proposed in this study to enhance the wrf model results. The ANN were trained 500 steps in Levenberg-Marquardt algorithm using the sign sets obtained by dividing the sign for 2 days training and 1 day test. The trainings were repeated. Logarithmic sigmoids were used in all the hidden layers of the network, and pure linear activation function is used in the output layer. WindSim Model, which is based on CFD-Computational Fluid Dynamics calculations, is also used to downscale the wind speed values simulated by WRF. WindSim solves 3D Reynolds Averaged Navier-Stokes equations. In general, studies show that WindSim has better performance than other models if the terrain of the interested domain is highly complex. In this study, WindSim is simulated by using the best results of WRF model (grid 15 – G15). The chosen domain used in WindSim model has 3.000.000 grid cells and 20 vertical grid levels. Modified option was selected as the turbulence model, height of the boundary layer was chosen as 20 m, and Wake Model 1 was also used for the wake model. The results obtained with multiple regression analysis are found to have error values close but higher than ANN results. The lowest error values were obtained by using wind speed from ANN 60m, wind direction, virtual temperature, east-west and north-south temperature gradient parameters. Results of G15 have better results compared to the work done by using parameters from the other grids (totally 16 grids) separately and gathered. It was seen that the parameters obtained from the levels of 700 hPa and 500 hPa were not very effective for ANN and better results were obtained with the data of 60 meters. The nRMSE value calculated with the wind speed from the G15 grid and the observed wind speed was 16.5% and with using ANN this error value was reduced to 12%.