Historical model forecast and observed visibility datasets

yyyy_spot[00/12]_wrf61_prev18.csv: 

This dataset includes visibility observations, station latitude and longitude, time and model forecast variables from 2014 to 2020.

The visibility observations were obtained from the Meteorological Information Comprehensive Analysis and Process System (MICAPS), which provided various weather measurements including temperature, pressure, humidity, precipitation, visibility, current weather, and other meteorological variables. Meteorological measurements were recorded at 3-hour intervals. Besides, the location and time of each record were provided.

The model forecast variables are from the Weather Research and Forecasting Nonhydrostatic Mesoscale Model (WRF-NMM) provided by the US National Weather Service (NCEP). The WRF-NMM produces hourly predictions for a 60-hour forecast period, launching twice daily at 00:00 UTC and 12:00 UTC. Forecast variables incorporated in the model encompass temperature, relative humidity, cloud cover, wind speed, precipitation, and more. This dataset includes 26 variables, which are listed below.

| Variables | Long Name | Description |
| ---- | ---- | ---- |
| DPT_GDS3_HTGL | Dew point | Height in 2 meters |
| HGT_GDS3_0DEG | Geopotential | height Level of 0 deg (C) isotherm |
| HGT_GDS3_CEIL | Geopotential | height Cloud ceiling |
| HGT_GDS3_HTFL | Geopotential | height Highest tropospheric freezing level |
| HGT_GDS3_SFC | Geopotential | height Ground/water surface |
| T_CDC_GDS3_EATM | Total cloud cover | Entire atmosphere |
| H_CDC_GDS3_HCY | High cloud cover | High cloud layer |
| M_CDC_GDS3_MCY | Mid cloud cover | Mid cloud layer |
| L_CDC_GDS3_MCY | Low cloud cover | Low cloud layer |
| PLI_GDS3_SPDY | Parcel lifted index (to 500 hPa) | Layer between two levels at specified pressure difference from ground to level |
| LFT_X_GDS3_ISBY | Surface lifted index | Layer between two isobaric levels |
| PRES_GDS3_SFC | Pressure | Ground/water surface |
| PRMSL_GDS3_MSL | Pressure reduced to Mean sea level | Mean sea level |
| MSLET_GDS3_MSL | Mean sea level pressure | Mean sea level |
| P_WAT_GDS3_EATM | Precipitable water | Entire atmosphere |
| POP_GDS3_SFC | Probability of precipitation | Ground/water surface |
| R_H_GDS3_HTGL | Relative humidity | Height in 2 meters |
| R_H_GDS3_HYBL | Relative humidity | Hybrid level |
| SPF_H_GDS3_HTGL | Specific humidity | Height in 2 meters |
| SPF_H_GDS3_SPDY | Specific humidity | Layer between two levels at specified pressure difference from ground to level |
| TMP_GDS3_HTGL | Temperature | Height in 2 meters |
| TMP_GDS3_SFC | Temperature | Ground/water surface |
| U_GRD_GDS3_ HTGL | U-component of wind | Height in 10 meters |
| U_GRD_GDS3_SPDY | U-component of wind | Layer between two levels at specified pressure difference from ground to level |
| V_GRD_GDS3_HTGL | V-component of wind | Height in 10 meters |
| V_GRD_GDS3_SPDY | V-component of wind | Layer between two levels at specified pressure difference from ground to level |

To integrate the gridded model forecast data with sparse observation data, we have a two-step pre-processing process. Firstly, we synchronized the time zones of two datasets to UTC+00:00. Secondly, we used the Inverse Distance Weighted (IDW) approach for spatial matching. As a result, we generated a multi-variable time series dataset as follows:
 \(D=\{X_{N \times M \times T}, Y_{N \times T}\}\) 
where \(X\) is a three-dimensional tensor of size \(N\times M\times T\) (corresponding to \(N\) samples of \(M\) forecast variables of \(T\) hours), and \(Y\) is the observational data of size \(N\times T\) (corresponding to \(N\) samples of \(T\) hours). A sample from dataset D is represented as \(s= \{ x_{M \times T}, y_T \} \in D\). In this datasets, \(M=27\), and \(T=60\). Besides, this dataset includes 18 hours of visibility observations before the model launching. It should be noted that some observations are missing for some reason.

yyyy_spot[00/12]_wrf6_prev18.csv: 

This datasets are from yyyy_spot00/12_wrf61_prev18.csv and is a sliding-window extraction of 61 forecast moments, with each sliding-window consisting of 6 hours.