1. Introduction
Wind forecasts play an important role in environmental prediction, where surface winds directly respond to atmospheric disturbances while also driving a variety of ocean effects, including gravity waves, surface currents, and sea level changes. In the renewable energy sector, the accuracy of such forecasts impacts the estimated energy production, leading to a series of economic and social consequences both in the short and long terms. According to the Global Wind Energy Council [
1] and the International Renewable Energy Agency 2021 [
2], Brazil holds the eighth position in the global ranking of wind energy, with 17 GW of installed capacity, coming from more than 750 wind farms and 10.000 wind turbines. Most of them (89%) are located in the northeast of Brazil, at tropical latitudes between 15° S to 0° N. Moreover, as reported by Associação Brasileira de Energia Eólica (ABEEólica, [
3]), Brazil has a wind power potential to reach 500 GW, which would attend three times the current country’s energy demand.
Vinhoza and Schaeffer [
4] assessed the wind energy potential in Brazil using a spatial approach, mapping the most economically attractive areas for offshore wind deployment. Santos et al. [
5] and Silva et al. [
6] investigated the wind energy utilization in Brazil in terms of its regulatory framework and national opportunities, describing the existence of a vast and reliable wind potential for immediate utilization in Brazil. Interestingly, Filgueiras and Silva [
7] and Lima and Mendes [
8] discuss that the overall power generation in Brazil is predominantly hydroelectric, which has led to an increasing vulnerability under changes in the global and local climate [
9], with impacts on the hydrological cycle; however, considering the seasonal cycle, the country reaches the highest wind speed exactly when the rate flow of the main rivers is low—raising the strategic importance of wind energy expansion, combined with high-quality environmental predictions. Therefore, wind forecasts in the northeast of Brazil have gathered a wide range of applications and demands.
Several high-quality wind forecast systems have been developed in recent years, using either numerical weather prediction models (NWP) or machine learning models. Jacondino et al. [
10] produced an hourly day-ahead wind speed forecast for two wind farms in northeast Brazil using the NWP model WRF (Weather Research and Forecasting; [
11]), where sensitivity experiments with different physical options were evaluated. Zucatelli et al. [
12] developed a recurrent neural network to perform 1 h wind forecasts in Brazil, following a recursive application for the subsequent hours. Samet et al. [
13] provided an evaluation of neural network-based methodologies for wind speed forecasting using multi-layer perceptron neural networks. Additional examples of artificial intelligence for wind forecasts worldwide are referred [
14,
15,
16,
17,
18,
19,
20,
21,
22,
23,
24]. However, all the aforementioned studies are focused on short-term forecasts, covering time scales of hours to a few days, whereas the mid-to-long term forecasts beyond 10 days remain uncovered and with a large demand. Even the recent study [
25], for example, which aimed at relative long-term predictions according to authors, was bounded to a 3 day forecast.
In this context, the present study proposes a new wind forecast system focused on forecast ranges from 1 to 46 days, with special attention to the subseasonal scale. Our goal was to combine the traditional NWP global models, where the coarse resolution does not capture regional features, with multilayer perceptron neural networks (MLP-NN, [
26]) trained with local measurements. Such a hybrid approach has been used by Xu et al. [
27], who argued that the current mainstream method for wind speed forecasting involves the combination of numerical meteorological models with statistical post-processing; nevertheless, their system is also restricted to 72 h forecasts, while our present study evolves this period to 46 days.
This initiative has been conceived for operational applications, i.e., the forecast system must be reliable, simple to maintain and to fix, with low computational cost, and robust. This paper starts with a description of the local wind characteristics in the northeast of Brazil (
Section 2.1), then it explains the numerical weather prediction using a global ensemble forecast and its importance for mid-to-long forecast ranges (
Section 2.2), followed by its evaluation in
Section 3.
Section 4 describes the neural network model for post-processing. The results of the hybrid model, combining the NWP model and the NN model, are presented in
Section 5 and discussed in
Section 6, including the operational application of the new prediction system. The conclusions and future plans are described in
Section 7. The research steps followed throughout this paper are summarized in the next
Figure 1.
Finally, it is import to highlight that wind, the horizontal component of air velocity, is a vector quantity that is described as complete with two numbers (scalars), namely velocity and direction. In order to cover the intensity and direction information, throughout this entire paper (observations, numerical modeling, and neural network modeling) we use the wind components (zonal) and (meridional). This is how the numerical models are integrated and how NNs consider the wind information. Circular or cyclic variables cannot be directly applied in NNs, otherwise similar directions, for example 1° and 359°, would be misinterpreted as being very distant to each other. Therefore, working with components and preserve the speed and directional information that can be later obtained, at the final visualization step, using wind speed as and the wind direction through the arc tangent function .
As a forecast mostly conceived for practical applications in the energy sector, the focus of this study was on wind speed, which is the most important input in the wind power equation:
where
is the air density,
A is the rotor area of the wind turbine, and
is the wind speed. Note that the wind power increases with the cube of the wind speed, i.e., doubling the wind speed gives eight times the wind power. This reinforces the accuracy of wind speed forecasts as being extremely important for wind energy applications.
3. Assessment of the Numerical Ensemble Forecast Data
An ECMWF archive of three years of ENS-Extended 46 day’s forecasts was selected for the domain, covering latitudes 40° S to 10° N and longitudes 60° W to 20° W, with a temporal resolution of 3 h. This is the NWP dataset selected for this paper and evaluated in this section. The ensemble point output for the positions of each of the 16 stations (
Figure 2) were extracted from this data archive, and paired with the observations (
Figure 5)—this allowed a complete assessment of the ensemble followed by the construction of post-processing neural network models. This ensemble dataset consisted of an array of 16 points containing three years of forecast cycles (twice a week, on Mondays and Thursdays), each one with 46 days of forecast lead times. By simply selecting one of these points, and plotting consecutive forecast cycles (x-axis) with the forecast lead time on the y-axis, it is possible to build a matrix containing the entire forecast for each point.
Figure 5A shows an example for a station in the state of Maranhão, where the direct difference (model minus measurement) is plotted in
Figure 5B.
The graphics of
Figure 5A are very rich, illustrating several features in the same plot. Firstly, it is possible to confirm the seasonal cycle with strong winds from August to November and light winds from February to June, proving that ECMWF simulations agree with observations for this time scale. Secondly, the effect of resolution is seen through the discontinuity in the wind intensity on the y-axis, where the first 15 days (0.2° of spatial resolution) have stronger winds than the forecast ranges from 16 to 46 days (0.4° of spatial resolution).
Figure 5B shows accurate forecast instants in white color, which are typically found during the first forecast days (lower y-axis values) close to the analysis (nowcast, y = 0), progressively growing at longer forecast lead time due to the chaotic component of the atmosphere. However, the “stripe” shape of
Figure 5B suggests that specific events of intensification and attenuation of wind speeds tend to loose predictability, showing poorer skill, and such lower model performance tends to remain similar towards the forecast time—i.e., when an oscillation in the wind intensity is not properly modeled, this tends to remain inaccurate throughout the whole forecast range. Such a characteristic revels two important points: (1) the numerical model and its resolution cannot properly simulate certain types of events and (2) the overestimation of winds tends to be followed by underestimation at scales of 10 to 50 days, and this error is not solved even at short forecast ranges. Lee et al. [
46] and Camp et al. [
47] discuss the challenges of subseasonal to seasonal (S2S) atmospheric forecasts in tropical areas.
In order to provide a complete assessment of the ECMWF ensemble forecast, four error metrics were calculated as a function of the forecast lead time at each station. Equations (2) to (5) describe the metrics, where
are the measurements and
the ECMWF forecast, and the overbar indicates the arithmetic mean. The systematic and scatter errors of the forecasts will be extensively discussed in this paper using the bias and scatter index, respectively. The scatter index can be interpreted as a percentage error when multiplied by 100. Typically, the pattern of lack of predictability at longer forecast leads is visualized by the increase of SI and the decrease of CC—the two most difficult metrics to improve. The next
Figure 6 and
Figure 7 show the results for each cluster, for wind speed and atmospheric pressure at mean sea level.
Figure 6 and
Figure 7 have common features that illustrate the general characteristics of the atmospheric modeling, such as the rapid increase of SI and decrease of CC in the first 15 forecast days. After this range, from 20 to 46 days, errors tend to be more stable and very large. Additionally, both clusters show better performances for the arithmetic ensemble mean (EM) of the ensemble members when compared against the individual members. This difference is less evident at the analysis (nowcast) but it grows as the forecast lead time increases, confirming the importance of the ensemble approach at mid- and long-forecast ranges. Such an effect is found in the SI and CC but not in the bias, where the error of EM and individual members are very similar. Therefore, the systematic errors are not reduced by the ensemble approach, which requires post-processing bias-correction tools.
In Maranhão, the SI of wind speed varies from 20 to 28% for the EM and from 24 to 38% for the single members. The CC at the same state varies from 0.83 to 0.65 for the EM and from 0.75 to 0.50 for the single ensemble members. The results highlight again the differences between the ensemble forecast and deterministic runs. In Piauí, there is more dispersion among the stations inside the same state, especially in the first 15 days. The SI of wind speed varies from 21 to 29% for the EM and from 24 to 34% for the single members. The CC varies from 0.8 to 0.63 for the EM and from 0.73 to 0.42 for the single ensemble members.
Regarding the systematic errors, note the differences between the two clusters, reflected in the bias plots of
Figure 6C and
Figure 7C. In Maranhão, bias is mostly positive, reaching 2 m/s within the first forecast days—which indicates that ECMWF forecast winds are more intense than measurements (model overestimation). However, in Piauí, where the winds have been shown to be more intense than in Maranhão, the bias plot has negative values, i.e., the observations are more intense than ECMWF forecast winds (model underestimation). This feature suggests that the NWP global simulations from ECMWF ENS-Extended tend to provide similar wind intensities in both states, while the observations prove they are significantly different, with Piauí containing stronger winds. Therefore, despite the quality of the ensemble forecast, the local characteristics and differences of surface winds in the northeast of Brazil are not properly simulated.
Another important limitation of the extended 46 days ensemble forecast can be found in the evolution of the error metrics throughout the forecast time, passing from day 15 to 16. The wind errors, in both states, show a very large jump at this boundary after 16 days, which reflects the different resolution of the model, from 0.2° to 0.4°—already found in
Figure 5. The higher resolution in the first 15 days provides stronger winds than the coarser resolution from days 16 to 46—clearly visualized in the bias plots. By analyzing the entire forecast range from 0–46 days, this problem becomes more critical, since the lower wind speed values, due to the model’s resolution, can be interpreted as a light wind period approaching. Interestingly, such strong discontinuity found in the wind errors is not present in the atmospheric pressure (hPa) error plots. The CC of pressure, for example, smoothly decays with increasing forecast days. The bias plots reinforce this feature, where the errors are nearly constant for the pressure. It is possible to conclude that the ECMWF ensemble forecast provides quality predictions of surface pressure, which responds to large-scale atmospheric phenomena, whereas the surface winds suffer from errors due to the influence of regional scales that require higher resolution and proper physics.
Therefore, for the post-processing algorithm to be attached to the output of the ensemble forecast, it must be able to bias correct the ensemble mean of the wind speed by learning the local characteristics from measurements, while still preserving the proper large scale prediction coming from the NWP model.
4. Neural Networks for Post-Processing
The neural network (NN) model selected for post-processing the ECMWF ensemble forecast was a Multilayer Perceptron Neural Network (MLP-NN, [
26]), which is a simple and powerful nonlinear mapper suitable for operational applications. A complete theoretical description is provided by Haykin [
48] and several applications in environmental sciences are found in Krasnopolsky [
49]. Samet et al. [
13] confirm that MLP-NNs are widely utilized in wind forecasting applications.
The MLP-NN is a feed-forward artificial NN that uses supervised learning, containing three or more layers: one input layer, one or more hidden layers, and one output layer. The next equation (6) presents the model, where
are the inputs,
the outputs, and
and
are the model parameters to be optimized. The total number of inputs and outputs are
and
, respectively. The first summation represents the “neuron” of the network, with the activation function hyperbolic tangent (
), where
is the total the number of neurons. Among the many activation functions available,
have shown very good results while still preserving the simplicity, as its derivative
.
The optimization of
and
was done with backpropagation training using gradient decent. In fact, it has been found that the stochastic gradient decent that evolved to the adaptive moment estimation (ADAM) described by Kingma and Ba [
50] provides better results and thus it was the method selected for this study. ADAM is especially suitable for sparse gradients and variables with strong noise components, which is the case of the environmental problem of this paper. Inspired by [
34,
51], the target variable of the MLP-NN model is not directly the wind speed but the residual, i.e., the difference between the ECMWF forecast and the observations—illustrated by
Figure 8. Therefore, the NWP modeling from ECMWF provided the ensemble mean as the first outcome that was directly compared and subtracted from the local measurements to build the residue signal. When the hybrid model is running, the ECMWF provides the EM that is added to the residue forecasted by the NN, to finally compose the forecast of wind speed. Due to the chaotic process of the atmosphere and lack of high-frequency predictability after two weeks of forecast [
52], the daily average was used in the hybrid model—which avoids the excess of noise that contaminates the model.
At this stage, it is important to discuss the construction of input and output dataset for training. The data structure is represented by
, described below, containing
variables coming from the ECMWF forecast (wind, temperature, humidity, and pressure) and
records (or quality-control measurements selected). In fact, the decision of
is more complex due to the two additional dimensions associated with the forecast lead time and several stations (positions) inside each cluster.
The NN prediction for each forecast lead time can be done independently, building one NN per forecast time, which would preserve the space but it would increase the number of NNs—leading to a cumbersome operational system. On the other hand, a new variable containing the information of forecast horizons (1 to 46 days, once daily averages have been used) can be added to the NN, concatenating the forecast dataset into a new column and increasing , related to the number of lines of . Hence, the new space becomes , where all the forecast times are appended into the NN and, therefore, one single NN model can predict different forecast leads—which has been the approach selected for this study.
Similarly, the spatial modeling for each cluster (Maranhão and Piauí) leads to three options in the NN:
- (i)
Build one NN for each station, independently. Therefore, the datasets of neighboring stations are not included into the NN training and modeling, i.e., the inputs (ECMWF forecast) and outputs (residue of wind) are restricted to that specific station.
- (ii)
Build a large NN where inputs are the information from all the stations inside the cluster (10 for Maranhão and 6 for Piauí) and the outputs are the residues of wind speed for the same stations. In this way, the dataset is expanded in , where corresponds to the environmental variables at each station. The data space for this approach, for example at Maranhão with 10 stations, would be .
- (iii)
Build one NN where inputs and outputs are variables for each station, which are one by one appended to generate a new dataset, where two new variables’ latitude and longitude are included to distinguish the stations. In this case, the training consists of an NN model passing through each station inside the cluster during the process. Note that the training data is constructed by appending the dataset of each station consecutively, so that the NN minimizes the error at the whole group of stations. Following the same example at Maranhão with 10 stations, the new space becomes
. It is important to highlight that in (ii), the number of input variables and dimension
is significantly augmented (10 times in Maranhão), whereas in (iii), the number of variables is just
, while the number of records is significantly expanded—following a similar approach of space/time trade proposed by the regional frequency analysis [
53].
There are pros and cons in each option described above. The first option (i) is the worst strategy because it ignores the neighboring data that could help improve the optimization, wasting the opportunity to increase the training dataset. The second option (ii) has the benefit of joining all the information inside the cluster. However, by appending the columns with several stations (increasing
), if one station has missing data, it compromises the entire record
, which means that only instants when all the stations have quality measurements could be used as training records. It is well-known that observations often suffer from outliers and wrong data that are excluded in the quality control process, as well as the initial/final dates of data collection can differ from one station to the other, thus limiting the final number of days used for training. Moreover, higher dimensions
usually require larger networks with more neurons, as described by Krasnopolsky [
49], which also demand more data for training. Therefore, expanding
while maintaining
is actually the opposite of what a good strategy should consider. Additionally, the NN structure of (ii) only allows the model to make predictions for the specific stations utilized for training, i.e., if one new station is included in the future, the model must be re-trained, and the new point can only be included when gathering enough duration to time-match with other stations and provide a sufficient coincident period of time for training.
The third approach (iii) combines several advantages and thus has been selected for the NN model implementation. The first and most important is the option of including all the stations for NN training no matter the duration of measurements or data gaps. By training the NN model at different stations with information of latitude and longitude, the NN performs a spatial mapping (position becomes a new dimension), so whenever new points are included in the future, the existing model can directly generate forecasts for these new positions. Furthermore, the new points with recent and short data collection can also be included in the NN re-training without having to modify the data structure or number of input/output variables. Therefore, each station bringing its measurements to the NN training contributes to the regional modeling by inputting short-scale information and improving the resolution at the region. Another important advantage of (iii) is that the NN preserves a small dimension (
) while increasing the data records
for training during the process of appending all the stations into a single dataset. In summary, latitude and longitude are two dimensions for spatial modeling, and time and forecast lead time are two additional dimensions as well. This data structure and training strategy was conceived by Dr. Vladimir Krasnopolsky in [
34], which allowed the successful construction of a global NN model in [
51].
In practical terms, two separated NNs were built and trained for each cluster (Maranhão and Piauí) and, due to the effect of predictability profile and model resolution on the wind forecast and intensity, it was decided to separate the 46 days of range into two slices of 1–15 days and 16–46 days. Therefore, four NNs were developed to cover two forecast ranges and two clusters. The last step was to include the methodology proposed by Krasnopolsky and Lin [
54], who argues that an ensemble of NNs outperforms single NN models alone. The NN ensemble construction can be easily implemented by managing the random initialization of
and
through the “seeds”. It is possible to generate several
-independent NNs with different random initial conditions and, as proven by [
15,
54], the arithmetic mean of results provides better prediction than using a single NN alone. Additionally, the spread of those
predictions provide information of uncertainty in the forecast, similar to what is traditionally evaluated in NWP models. Moreover, by generating a large number of NNs, it is possible to exclude the networks with poorer performance and pick only the networks with the best results. This methodology has been implemented for the NN modeling so, in terms of outputs, the schematic of
Figure 8 can be expressed as:
where EM is the ECMWF ensemble mean and
is the arithmetic mean of
networks with
parameters
. The result of the hybrid system,
, is the final wind forecast.
Expanding the discussion of dimensionality of the NN model and training data, but now focused on the input variables with meteorological information, it is crucial to investigate how the ECMWF outputs will feed the NNs. Being an ensemble with 51 members (50 perturbed members plus the control) and considering the variables wind (
and
components), pressure, temperature, and humidity, the number of environmental inputs for the NN reached 255. Zaki and Meira [
55] reinforce the challenges and possible problems involving high dimension, which is especially critical in operational applications that require reliable and robust models. Thus, a feature selection was conducted to: (i) analyze the importance of each component of meteorological information (wind, pressure, temperature, and humidity) and how this inclusion/exclusion affects the performance of the model, first in the input and then in the output;(ii) analyze if the 51 members are necessary as inputs or if it is viable to replace them by their probabilistic moments (mean, variance, skewness, and kurtosis), such as applied in [
34,
51]. It was found that including all the 51 members significantly increases the number of inputs which demands larger NNs, which lead to training difficulties. In practical terms, the performance in the training set was satisfactory while in the test set it was compromised. Consequently, it was safer to move forward with the probabilistic moments, since these captured the characteristics of the 51 members while reducing the number of variables.
The next
Figure 9 illustrates part of the feature selection process, where the importance of each probabilistic moment from each meteorological variable relates to the target variable (residue of wind speed, in black in the figure). The feature analysis takes into account the Pearson’s Correlation and the Predictive Power Score. Besides the straightforward probabilistic moments, the L-moments [
56] were included, due to being less affected by outliers in the time-series.
Figure 9 shows a nearly equal importance of temperature, humidity, and pressure into the outputs. Regarding the inputs, the first moment (mean) shows a greater importance of humidity and temperature, and lower importance of atmospheric pressure. This pattern completely changes for the second moment (variance), where humidity loses its relevance. Moving to higher moments, the impact of variables drops significantly but the zonal (
) component of the wind remains important. Nonetheless, when the skewness of the ensemble members is different than zero (asymmetry), it can provide valuable information, once it directly affects the ensemble mean—a common effect at longer forecast leads.
Figure 9 and later results indicate that L-moments do not contribute to the NN modeling when compared to the standard probabilistic moments.
The feature selection algorithm initially applied a methodology inspired by [
57,
58], followed by the analysis of all the combinations of input and output variables, evaluated in both clusters and through the 46 days of forecast range, considering the performance in the training and test/validation sets. The best combination of variables gathered for inputs was the ensemble mean (temperature, humidity, wind speed), variance (temperature, and
and
wind components), and skewness (humidity, and
and
wind components). For outputs, this was the wind speed, accompanied by the meridional component of the wind only. This gave 10 environmental variables for the inputs, which also included the forecast time, sine and cosine of time (to properly identify the time of the year and differentiate seasons), and latitude and longitude—a total of 15 inputs. The number of outputs was two.
The NN training as well as other tests involving feature selection, optimization of the number of layers and neurons (next section), and evaluation of final results were carefully conducted with a proper data division into training set, test set, and validation set. The most basic requisite was to train the NN model using records that were not selected for testing. The leave-one-out cross validation method with three cycles (
Figure 10A) was implemented, where two thirds of the dataset was used for training and one third for testing the network—with intercalated indexes that can be joined to reconstruct a test time-series with the exact same time as the original one. However, such a method carries a limitation of building training and test sets within the same time interval, i.e., both sets correspond to the same atmospheric conditions inside an ergodic system.
It is extremely important to evaluate the NN considering a different time interval and atmospheric conditions, in order to evaluate the extrapolation and generalization of the model—which is critical for operational implementation. Therefore, a validation set was built (
Figure 10B), where the NN was trained and tested using two years of data, and validated using a separated year. This strategy allowed for the simulation of a real forecast in operation, for when the NN cannot be constantly trained. In reality, the validation set evaluates the extreme scenario of not re-training the NN model for an entire year—which is not necessarily the real case with the NN being re-trained at least twice a year. The total number of records in this three-year period was 46,656, which were used for the model optimization and evaluation as explained. Therefore, the total input dimension was 46,656 × 15 and the output dimension was 46,656 × 2.
6. Final Discussion
Results of
Figure 13,
Figure 14 and
Figure 15 summarize a NN forecast product developed through multiple steps, including the initial quality control and clustering of stations, construction of a dataset for training, dimensional analysis, development of a proper NN architecture and training strategy, and development of an ensemble of NNs. In fact, the final forecast consisted of a hybrid product (
Figure 8 and Equation (8)), joining the NWP from ECMWF and the post-processing NN model. This part is important to be emphasized because the large-scale and mid-to-long range atmospheric modeling is mostly performed by the global ensemble forecast of ECMWF, with the NN in charge of bias-correction and regionalization of this initial prediction. The NN model construction could preserve the simplicity of the MLP-NN, developed by Rumelhart et al. [
26], with a theoretical description by Haykin [
48], and environmental application at Krasnopolsky [
49], which has been fundamental to the success of the tests involving the NN architecture and data structure [
55], and the NN ensemble [
54].
One of the most critical problems in the EMCWF prediction, discussed in this paper, is the severe underestimation of wind speeds in the state of Piauí. This problem was solved with the NN ensemble, with improved predictions for the whole forecast range from 1 to 46 days. Besides
Figure 14 and
Figure 15 showing the bias,
Figure 16 presents a real-case simulation (reforecast) for a station in the state of Piauí, comparing the initial ECMWF forecast with the results from the post-processing NN. The underestimated ECMWF winds between August to December are visualized with yellow colors in the Figure (winds around 9 m/s), while the bias-corrected winds from the NNs are seen with orange colors (winds around 11 m/s). As a non-linear mapper, the NNs do not uplift the wind events equally or with constant rate, so the model captures the local weather patterns in the given environmental input variables, to properly bias-correct the ECMWF forecast when necessary.
The success of the NN post-processing model goes beyond the bias correction, as pointed out by
Figure 14 and
Figure 15. The RMSE could be reduced and CC increased, with the performance of test and validation sets ensuring the good generalization and robustness of the NN ensemble. The state of Maranhão had the best increase of CC while the state of Piauí had the best RMSE decrease—which reflects the small differences in the local climate in both states. The state of Piauí has less influence of the coastal orientation, orography, and consequently, the regional effects—having a more direct response to large-scale meteorological events, modeled by the global ECMWF ensemble. Such a prediction, as presented, has systematic errors that could be properly corrected by the NN model; however, the oscillation patterns, which influence the correlation coefficient, are mostly affected by these large-scale global phenomena as well, and so the NN as a local model could not map this part of the signal. A larger space-time machine learning model, covering a broad domain, could potentially capture this part of the forecast in Piauí. In Maranhão, however, the regional influence is more pronounced, so the CC could be better improved by the NN model; the bias, however, is not as high as in Piauí, so the contribution from the NN is relatively low. Since RMSE is a composition of scatter and systematic errors [
60], the percentage of decrease of RMSE in Maranhão is also lower than the same in Piauí. Nevertheless, the NN results in Maranhão for the 15-day forecast have the same RMSE as the ECMWF forecast for 8 days, i.e., the NN post-processing represents a gain of one week in predictability.
Considering the absolute differences of error metrics between the NN results and ECMWF forecasts, the mostly improved forecast would be the longer range from 16 to 46 days. Nevertheless, even using the hybrid modeling proposed in this paper, this range remains a great challenge. This can be visualized through the differences between the NN test set and validation set. While in the range of 1–15 days, both sets have very similar values; the results completely change for the 16–46 days range, where the test set shows better values than the validation set. Such a difference suggests that the NN lacks generalization in the long-range when compared to the short-range forecast. This feature is more pronounced in the CC and RMSE, and not in the bias plots, so the NN model remains a good bias-correction tool with proper generalization even in the long-range. The impact seen in the CC and RMSE for long ranges reflects the general lack of predictability and accumulated chaos in the atmosphere that has been a meteorological challenge, especially in tropical regions.
Although a direct comparison of forecasting performances with previous studies is not possible, due to different locations and forecast ranges, it is still important to contextualize our results and method with other references and algorithms. For the short-range segment shown in
Figure 14, the RMSE in the present study varied from 0.8 to 1.0 m/s and bias was between −0.1 and 0.1 m/s. The wind forecast obtained directly using the numerical model WRF described in [
10] resulted in an RMSE between 1.5 and 2.1 m/s within the 48 h range. Using wavelet decomposition and recurrent neural network, [
12] we obtained an RMSE from 0.76 to 2.03 m/s. At very short-term forecasts using 10 and 5 min wind sample intervals, [
13] this produced an RMSE from 0.03 to 0.76 m/s, with the best performance obtained with wavelet transformation of MLP models using two training algorithms, Levenberg–Marquardt and Bayesian Regularization. Authors in [
14] have shown RMSE values from 0.07 to 1.56 m/s with the best results from a new hybrid decomposition technique (HDT) combined with an ensemble empirical mode decomposition adaptive noise (CEEMDAN) and empirical wavelet transform (EWT). Authors in [
15] have reached RMSE values from 0.52 to 0.89 m/s and the best performance with a wavelet analysis technique combined with negative correlation learning neural network and an ensemble structure optimized using particle-swarm optimization WAT-NCL-PSO. Authors in [
16] obtained RMSE from 0.33 to 1.15 m/s using an NN model integrated with a modified firefly algorithm and particle swarm optimization. Authors in [
17] obtained RMSE from 0.014 to 1.32 m/s and reported the best results using a wavelet packet decomposition, density-based spatial clustering of applications with noise, and the Elman neural network (WPD-DBSCAN-ENN), and [
18] obtained RMSE from 0.26 to 0.89 m/s with the best results using singular spectrum analysis and general regression neural network with CG-BA (SSA-CG-BA-GRNN), where CG-BA is a modified bat algorithm (BA) with the conjugate gradient (CG) method developed to optimize the initial weights. These studies suggest that there is a wide range of algorithms with great potential to further improve our wind forecast, moving to more complex models and signal decomposition. Furthermore, instead of searching for one single method, [
20] successfully combined a data preprocessing technique with a linear combination of multiple prediction models to produce their forecast—which is a promising multi-model ensemble approach to be implemented in the next steps of this present study.