E the NN is often a scalar function because it always outputs a single worth (i.e., either Tmax or Tmin ). Considering the fact that we wanted to illustrate the NN behavior we limited the amount of input parameters to two–this enabled us to visually show the behavior in the NN as 2D contour graph. We aimed to attempt out several PHA-543613 Technical Information setups of simplistic NNs, with different degrees of complexity and see how it affects the resulting behavior. We focused around the same-day forecast (forecast for precisely the same day because the radiosonde measurement was created). We wanted to make use of two profile-based input parameters that would generate a reasonably good forecast of either Tmax or Tmin . We experimented with different parameters derived in the vertical profiles. Ultimately, we chose the average temperature in the lowest layer in between the ground and 1 km and also the 90th UCB-5307 In stock percentile of RH in the layer in between the ground and 12 km (both parameters have been calculated from the information of your original profiles, without the need of interpolation to standard altitudes). The first parameter reflects the general temperature conditions within the boundary layer, that will rely on the season and also the general weather circumstance (the strong link among Tmax along with the temperature in the boundary layer is also clearly visible in Figure 2). The second parameter is often related with all the existence of cloudiness. As currently talked about, the clouds will weaken downward shortwave radiation close to the ground during the day, which reduces the temperature close to the surface. The radiosonde does not straight measure the existence of clouds. Still, it could be approximately inferred in the RH measurements (an RH value bigger than 90 indicates a high likelihood of clouds at that altitude). Apart from the possibility of either possessing none or at the very least some clouds, the cloud thickness also influences the downward shortwave radiation. If you will discover no clouds, the 90th percentile of RH will have a fairly low worth (i.e., substantially smaller sized than 100 ), whereas if a sufficiently thick cloud layer is present, the 90th percentile of RH will be close to one hundred . The analyzed NN setups are described in Table 1. We began using the most uncomplicated NN with only a single neuron (Setup A). We initially attempted working with the rectified linear activation function (ReLU), which did not perform effectively. The cause was that for the duration of training, the two weights and also the bias had been oftentimes set to adverse values, following which the training could not proceed any longer (this dilemma is referred to as the “dying ReLU” inside the literature). Exactly the same problem also occurred for other setups shown in Table 1, though not as often. The dying ReLU trouble could be avoided making use of a slightly modified version of ReLU named the Leaky ReLU, which includes a tiny slope for damaging values that enables the training to proceed even if the weight and bias have damaging values.Appl. Sci. 2021, 11,7 ofTable 1. Description with the simplistic neural networks consisting of only a few neurons. All setups utilised the exact same two input parameters, the typical temperature in the lowest layer in between the ground and 1 km plus the 90th percentile of RH within the layer among the ground and 12 km. The second column denotes the number of neurons in consecutive layers: input layer generally consists of two neurons for two input parameters and will not be incorporated inside the table, whereas the output layer always contains a single neuron. Leaky ReLU was applied as activation function for all layers in all setups. The shown MAE values represent the error of your sa.