Accurate surface ultraviolet radiation forecasting for clinical applications with deep neural network

Abstract

Exposure to appropriate doses of UV radiation provides enormously health and medical treatment benefits including psoriasis. Typical hospital-based phototherapy cabinets contain a bunch of artificial lamps, either broad-band (main emission spectrum 280-360 nm, maximum 320 nm), or narrow-band UV B irradiation (main emission spectrum 310-315 nm, maximum 311 nm). For patients who cannot access phototherapy centers, sunbathing, or heliotherapy, can be a safe and effective treatment alternative. However, as sunlight contains the full range of UV radiation (290-400 nm), careful sunbathing supervised by photodermatologist based on accurate UV radiation forecast is vital to minimize potential adverse effects. Here, using 10-year UV radiation data collected at Nakhon Pathom, Thailand, we developed a deep learning model for UV radiation prediction which achieves around 10% error for 24-h forecast and 13-16% error for 7-day up to 4-week forecast. Our approach can be extended to UV data from different geographical regions as well as various biological action spectra. This will become one of the key tools for developing national heliotherapy protocol in Thailand. Our model has been made available at https://github.com/cmb-chula/SurfUVNet .

Figure 1

Characteristics of UV and weather conditions at Nakhon Pathom, Thailand. Daily maximums are shown for UV irradiance, total ozone column, and AOD500. Daily averages are shown for cloud coverage. Dark lines indicate the average across 2009–2017. Shaded areas indicate the ± 1 standard deviation range. (a) Annual surface UV irradiance. (b) Annual cloud coverage. (c) Annual total ozone column. (d) Annual AOD500. (e) The distribution of cloud coverage in the validation set (UV data from year 2018). Both Silpakorn University’s observations and ERA5 data were shown. (f) The distribution of cloud coverage in test set (UV data from year 2019). Information from Silpakorn University is unavailable.

Figure 2

Schematic of SurfUVNet. (a) The underlying encoder–decoder neural network architecture showing the flow of data from the encoder to the decoder via the central connection denoted by S_t. LSTM and Dense indicates the Long Short-Term Memory and fully connected neural network layers, respectively. UV data from days prior to the forecast date are fed into the encoder part while UV data from the same date of previous year are fed into the decoder part. The model forecasts next-day UV radiation at 10-min resolution. (b) The auto-recursive mode for long-term UV forecasting. To forecast UV radiation for the next N days, SurfUVNet first forecast next-day’s UV radiation profile and then uses the prediction as input to forecast UV radiation profile for the day after. This process is repeated until the forecasts for the next N days are generated.

Figure 3

SurfUVNet accurately forecast antipsoriatic irradiance throughout the day. Results on Nakhon Pathom dataset were shown. (a) Comparison of the mean absolute percentage errors (MAPE) for the next-day antipsoriatic irradiance forecast between SurfUVNet (Seq2Seq-14) and four benchmark models (see “Methods” section). Previous day model simply predicts next-day’s UV radiation to be the same as today’s. Regression model refers to the regression model based on Earth–Sun distance and total ozone column currently in used by the Thai Meteorological Department. BiGRU is an artificial neural network architecture that is often utilized for time series forecasting. CNN-LSTM, and CNN-LSTM-SG are artificial neural network models that were recently applied to UV forecasting in the energy domain. The tags − 7, − 14, and − 21 designate the length of UV data, in days prior to the forecast date, that were input into each model. (b) Distribution of MAPE for the validation set (UV data from 2018) throughout the times of the day. Results for the best performing models, namely CNN-LSTM-SG-7 and SurfUVNet (Seq2Seq-14), are shown. (c) A similar plot showing distribution of MAPE for the test set (UV data from 2019). (d) Comparison of ground truth UV data and forecasts made by SurfUVNet for the validation set (UV data from 2018). Error bars indicate one-standard deviation ranges. (e) A similar plot for the test set (UV data from 2019).

Figure 4

SurfUVNet’s forecast error weakly correlates with cloud coverage. Violin plots showing the distribution of SurfUVNet’s forecast error in 1-h interval with various cloud coverage. Errors on the test sets (UV data from year 2019) are shown. (a) Nakhon Pathom dataset. (b) Tokyo dataset. (c) London dataset. (d) Heliotherapy sunbathing sessions planned by photodermatologist at King Chulalongkorn Memorial Hospital. Each data point that constitutes the violin plots correspond to the error between predicted and actual antipsoriatic irradiances that a patient would be exposed to if he or she were to sunbath according to dermatologist’s planning.

Figure 5

Long-term antipsoriatic irradiance forecasting. Results on Nakhon Pathom dataset were shown. (a) Diagram of two approaches for making long-term forecast: developing specific artificial neural network model for making forecast for a specific day that is a certain number of days into the future (left) and autoregressively using the next-day forecast as input for making forecast for the day after that (right). (b) Long-term antipsoriatic irradiance forecasting performance for up to 28 days into the future on the validation set (UV data from 2018) and the test set (UV data from 2019). Performance for SurfUVNet, the regression model based on Earth–Sun distance and total ozone column, and the best CNN-LSTM models were shown.