Results obtained from an Automatic Cloud Nowcasting System based on all-sky images
1Sieltec Canarias S.L. Santa Cruz de Tenerife, Canary Islands, Spain
2Galotecnia Redes Sistemas y Servicios SL . Santa Cruz de Tenerife, Canary Islands, Spain
3Dept. Física de la Terra i Termodinámica, Facultat de Física, Universitat de Valéncia. 46100 Burjassot (Valencia). Spain.
4Izaña Atmospheric Research Center (IARC), Meteorological State Agency of Spain (AEMET), Santa Cruz de Tenerife, 38001, Spain
Sieltec Canarias S.L. has developed a nowcasting algorithm for clouds, based on hemispherical sky images obtained by its automatic cloud observation system (SONA). SONA was a joint project with the Izaña Atmospheric Research Centre (State Meteorological Agency of Spain, AEMET). This algorithm analyzes cloud relative flow from all-sky images, allowing predicting when the Sun will be blocked by clouds in the near future, and thus when a sharp drop/increase in Direct Normal Irradiance (DNI) is expected.
In this work we show results of our cloud nowcasting algorithm evaluation by assessing randomly selected image sets in test stations equipped with all-sky cameras. In order to evaluate our cloud nowcasting algorithm, we used images of cloud cover predicted by our algorithm, versus the actual images of cloud cover.
Two places have been studied for testing the accuracy of the nowcasting system. San Cristobal de La Laguna (Tenerife, Spain, lat: 28.486, lon: -16.327) and Valencia (Spain, lat: 39.507, lon: -0.42). See online demo of the system on:
http://demo1.sieltec.es User: guest Password: guest
Cloud nowcasting procedure
The system studies the relative cloud motion between High Dynamic Range images (HDRI) in different temporal scales to detect cloud progress. This parameter will be the reference to determine which part of the clouds are going to block the Sun. The algorithm finds the remaining time until this part will reach the Sun. The accuracy of this temporal window depends on atmospheric conditions such as cloud stability, cloud layering, cloud’s altitude, orography of the all-sky camera’s site, etc.
The system works with different “sky states”:
- State 0. When there are not clouds, or they are far enough and moving very slow (the system is stable)
- State 1, appears when there are clouds approaching to the Sun and the previous situation was state 0
- State 2. Sun is blocked
- State 3. The Sun is going to be blocked, but the previous state was state 2; statistically this situation has more instability (cloud creation/destruction) than state 1 and the nowcasting uncertainty grows.
(A) Nowcasting process at state 0, the sky is basically clear or there are clouds but they are moving very slow (the system is stable).
(B) State 1. Clouds are moving towards the Sun. The previous state was state 0.
(C) State 2. Sun is blocked.
(D) State 3. There are clouds approaching to the Sun and previously the Sun was blocked (state 2).
Randomly selected validation sets longer than 30000 HDRI were used for this proposal. In Fig. 3 the nowcasting success degrees at different time intervals in Valencia and Tenerife are shown. Logically, the longer the prediction is, the less accurate it will be. Even so, we have for all nowcasting values at Valencia and Tenerife, over 25 minutes 84 and 85% respectively of success. Figure 3B1-3B2 tells us the accuraccy of nowcasting values over an specific time. In figure 3, it is also observed that if the nowcasting times are calculated from state 1 or state 3, state 1 show a better behaviour since corresponds to a stable sky situation.
Fig 3. (A1-A2) Nowcasting reliability in Valencia and Tenerife by intervals. (B1-B2) Nowcasting reliability in Valencia and Tenerife over specific times. In both cases, it is shown the results from the whole set of random validation sets, state 1 and state 3 situations.
When the system fails, it is defined an average time error as the average time by interval which the algorithm has overestimated the time until the Sun is going to be blocked (Fig. 4). That would yield to subtract a safe time margin to the nowcasting data, decreasing the temporal nowcasting time, just to be sure that the Sun is not going to be blocked before the time frame in any case. To sum up, the system is able to get a trustable nowcasting time and is flexible: the restrictions imposed by the users can make these times shorter or longer. Valencia has different orographic conditions than Tenerife, which affects the quality of the results, especially in San Cristobal de La Laguna the site has a complicated cloudy system, but attending to the results the system is trustable and useful in both places.
Fig 4. Average time error (A) Nowcasting reliability in Valencia. (B) Nowcasting reliability in Tenerife. In both cases, it is shown the results from the whole set of random validation sets, state 1 and state 3 situations.
Once the cloud nowcasting is working, adding to our system Direct Normal Irradiance data, we can predict the DNI when the Sun is blocked by clouds. We only have to study the past temporal series of DNI data, using non-linear neural networks and mathematical models to achieve a value of radiation for the near future. Next step will be to study the cloudiness shadow map over the terrain, and predict a forecasting radiation map based on it.
Our acknowledgement to CIAI’s Basic Systems Unit (Ramón Ramos, Rubén del Campo, Cándida Hernández, Virgilio Carreño, Fernando de Ory, Enrique Reyes, Antonio Cruz, Rocío López and Néstor Castro) and the staff of Dept. Física de la Terra i Termodinámica, Facultat de Física, Universitat de Valéncia. Also thanks to Sieltec Staff (Priscila Ramos, José Felipe, Hao Lu Higuchi, Tomás Martín, Enrique Pérez, Javier García, Ines Duranza, Antonio Dorta and Ramón Negrillo).
This work has been funded by the Ministry of Economy and Competititiveness (MINECO) and the European Regional Developments Funds (EDRF) throught project CGL2015-64785-R and CGL2015-70432=R, and the Generalitat Valenciana through the project PROMETEUII/2014/058.