Several deep learning models were applied in this study to propose a sufficient technique for multistep UV index forecasting. The developed convolutional neural network integrated with long short-term memory network (CLSTM) model was implemented and evaluated against various other techniques, namely long short-term memory network (LSTM), convolutional neural network (CNN), deep neural network (DNN), multilayer perceptron (MLP), extreme learning machine (ELM), random forest regression (RFR), extreme gradient boosting (XGB), and Pro6UV Deterministic model. These models were developed to predict three-time steps (10-min, 30-min, and 60-min) of UV index by employing various hyperparameters as shown in Table 3.
An evaluation study, which used several evaluation metrics, has revealed that the proposed model, i.e. CLSTM, found to have an accurate forecasting skill comparing to other competing models for all periods using all metrics (Tables 4 and 5). The presented results in these table confirm that the proposed model achieved better performance than other examined techniques. For 10-min horizon forecasting, results of Table 4 showed that CLSTM model achieved values of the evaluating metrics of 0.3817, 0.1887, 8.0086, 4.6172, and 3.9586 of \(RMSE, MAE , RRMSE \left(\text{%}\right), MAPE \left(\text{%}\right), and APB\) while the closest model was LSTM with values of 0.3936, 0.1939, 8.2569, 4.7709, and 4.0680, respectively. The same order can be noticed from the same table for 30-min and 60-min forecasting horizons where CLSTM was the best model followed by LSTM. Moreover, looking at results of other evaluation metrics, which are shown in Table 5, revealed that the current developed technique outperformed other competing models. The CLSTM model showed the best values for 10-min period in according to the values of those metrics with \(r=0.9964, WI=0.9907 , {E}_{NS}=0.9855, LM=0.9276, and KGE=0.9936\). On the other hand, the closest model was LSTM where it gave the following values: \(r=0.9961, WI=0.9901 , {E}_{NS}=0.9846, LM=0.9256, and KGE=0.9917\). Similarly, CLSTM had the best performance for other horizons, i.e. 30-min and 60-min.
As mentioned earlier, this study has developed a variety of models, including some machine learning techniques. Therefore, it is necessary to compare the performance of CLSTM, as a deep learning model, with the competing machine learning models. From Tables 4 and 5, it is clear that ELM was the best among other machine learning models, i.e. \(RFR and XGD\). It can be seen clearly that CLSTM achieved better RMSE values for all horizons. It had a value of 0.3817 for 10-min, 0.4866 for 30-min, and 0.5146 for 60-min while ELM had the following values: 0.4620 for 10-min, 0.5505 for 30-min, and 0.5996 for 60-min. Correspondingly, a quick look at Table 5 that shows that CLSTM presented better results than ELM. According to \(r\) values for instance, the CLSTM model gained 0.9964, 0.9940, and 0.9932 for 10-min, 30-min, and 60-min, respectively whereas the values of ELM were 0.9946, 0.9923, and 0.9907 for 10-min, 30-min, and 60-min, respectively. In addition, comparing with deterministic method, we can notice that this model achieved values of \(RMSE\) of 2.5319, 0.8800, and 0.8747, and values of \(r\) of 0.8534, 0.8613, and 0.8511 for 10-min, 30-min, and 60-min which are significantly far from the results of CLSTM model.
It is also important to notice that the current paper has made a significant contribution regarding UV index forecasting using an advance data-driven approach by considering various time steps. In addition, the present study has conducted an in-depth comparison between some of the most popular deep learning models, machine learning techniques, and a deterministic model. Addressing these two areas has led to contribute significantly to earlier studies that dealt with suggesting forecasting models for UV index, such as [9, 10].
Comparing to the closest works in the state-of-art, it can be noticed that the current study has produced a more effective forecasting model for UV index. This paper can be considered as more comprehensive than others. [10] examined six techniques, including four deep learning models and two machine learning models. The authors of this article used only three evaluation metrics, namely \(r, MAE, and RME\). Comparing with our results, it can be found that our proposed model achieved better performance in 10-min, 30-min and 60-min where it gave values of \(r\) of 0.9964, 0.9940 and 0.9932 for the above-mentioned time steps, respectively. [9] considered only machine learning methods, that is ELM, MARS and M5 Model Tree, and deterministic (Pro6UV) model. Moreover, they only utilized one-time period, which is 10-min, to forecast UVI.
Furthermore, the limitations of this work can be categorized as: 1) no other variables, such as climate measures, have been considered in developing CLSTM, 2) datasets were only acquired from one region, that is a city in Brisbane, QLD, Australia. As a result, some improvements can be done as possible future works, such as employing weather data as inputs when developing the model, using UV index data from different regions to examine the quality of the suggested model, and utilizing CLSTM in forecasting other phenomena's.