NEURAL achievements Ivica Vilibić, PI or 1001 reasons why UKF financed this project ;-) Ivica Vilibić, PI Institute of Oceanography and Fisheries, Split, Croatia

Project factography Project Title: Interpreting and forecasting Adriatic surface currents by an artificial brain (NEURAL) Principal Investigator: Ivica Vilibić Contracting Institution: Institute of Oceanography and Fisheries, Croatia Participating institutions: University of Ljubljana, Meteorological and Hydrological Service, Istituto Nazionale di Oceanografia e di Geofisica Sperimentale - OGS Funding Agency: Unity Through Knowledge Fund Project Period: 15 October 2013 – 14 October 2015 Project Team: I. Vilibić (IOF), Nedjeljka Žagar (UniLj), Simone Cosoli (OGS), Vlado Dadić (IOF), Damir Ivanković (IOF), Blaž Jesenko (UniLj), Hrvoje Kalinić (IOF), Hrvoje Mihanović (IOF), Martina Tudor (DHMZ), Jadranka Šepić (IOF) Individual consultant: Ivica Janeković (IRB) Web: http://jadran.izor.hr/neural

NEURAL project The main objective of the NEURAL project is to research and to build an efficient and reliable prototype of ocean surface current forecasting system, based on HF radar measurements, operational mesoscale weather forecast and neural network algorithms (Self-Organising Maps). 2 3 1 2

NEURAL project Project is composed of 9 tasks: Task 1. Setup of operational procedures for the HF radar network. Task 2. High-level quality control of new and historic HF radar data. Task 3. Mesoscale reanalysis and forecast of the atmospheric fields by Aladin/HR and WRF-ARW models. Task 4. Tuning the neural network parameters for optimal performance. Task 5. Obtaining characteristic patterns of surface currents and meteo model data through learning process. Task 6. Assessing the abilities and skill performance of the system. Task 7. Setting up procedures and protocols for operational ocean forecast. Task 8. Creating dynamic web content with forecast of ocean currents. Task 9. Dissemination of project results. Three project components: Component 1. Technological component – purchase, installation and operational maintenance of HF radars in the middle Adriatic, setting up operative WRF Component 2. Research component – historical and new HF radar data analyses, connection to reanalysis and operation winds by using Self- Organizing Maps (SOM) method, testing different SOM and NWP configurations and pre-processing algorithms, ... Component 3. Operational component – setting up automatic procedures, creation of operative SOM-based forecasts, development of web-oriented products, ... 4

Technological component Installation of HF radars in the middle Adriatic (jointly with HAZADR project): Two WERA shore-based remote sensing system used to monitor ocean surface currents, waves and wind direction. Mid-Adriatic WERA works at ~25 MHz. Installation in January/February 2014. Data availability: March 2014 – January 2015, May 2015-present. Spatial resolution: 1.5 km; Temporal resolution: 30 min. Ražanj (Brač) 5

Installation of HF radar at Cape Stončica (Island of Vis) Technological component Installation of HF radar at Cape Stončica (Island of Vis) 6

Technological component Temporal coverage Extreme libeccio (SW wind), 30-01-2015 Ražanj HF radar heavily damaged Spatial coverage of quality-controlled data Problems at the SW corner of the domain, probably result of Italian emitting sources at similar frequencies 7

Technological component Set up of WRF/ECMWF operative forecast for the Adriatic (FMF-UniLj): ARW, version 3.2 Horizontal resolution of 9 km (91x91 ) through 3 km (121x121) to 1 km (181x163) over the northern Adriatic 91 levels in vertical RRTM longwave & Dudhia shortwave radiation, WSM-3 microphysics, 5-layer thermal diffusion surface layer, the Yonsei University boundary layer scheme and the Grell convection parameterization http://meteo.fmf.uni-lj.si/NEURAL 8

Research component 9

Operational component Scheme of an operational coastal forecasting system based on the SOM method, which uses HF radar data and NWP mesoscale model executed in reanalysis mode for training, while forecasts are executed by using NWP model forecasts only.

Operational component HF radar data flow: Operational component Antenna hardware Mac + Codar software GPRS, EDGE, UMTS router lighthouse Internet MEDAS Application server Automatic FTP sincronization and database ORACLE database ORACLE application server web IOF Data Center Data Centre – network schema

Project was composed of 9 tasks: Task 1. Setup of operational procedures for the HF radar network. How to get the raw data in real time? How to process the data?  real time quality control procedures (min-max, spikes, neighbouring values, ...)  real time estimation of total current vectors from radials  data archiving  web oriented presentation in collaboration with HAZADR project Milestone 1. Operational procedures for HF radar network ready to use (month 4).  done in month 5 (March 2014), thanks to Damir and Vlado

Task 1. Setup of operational procedures for the HF radar network. WERA shore based remote sensing system used to monitor ocean surface currents, waves and wind direction. Mid-Adriatic WERA works at ~25 MHz. Availability: March 2014 – January 2015, May 2015-present. Spatial resolution: 1.5 km; Temporal resolution: 30 min. 13

Task 2. High-level quality control of new and historic HF radar data. How to get high-quality data for research? How to process the data in archives?  delayed quality control procedures  which methods? Signal to noise ratio, spatial comparison, interpolation, ... Re-assessment of already measured values (SeaSonde, north Adriatic) Assessment of newly achieved data (WERA system, middle Adriatic) in collaboration with HAZADR project Milestone 2. Historical hourly HF radar data from the northern Adriatic re-assessed through delayed-mode quality control procedures (month 6).  done in month 4 (February 2014), thanks to Simone and Hrvoje M. Milestone 3. HF radar data from the operational network in the middle Adriatic assessed through automatic delayed-mode quality control procedures (months 12-24).  done in month 23 (September 2015), thanks to Hrvoje M.

Task 2. High-level quality control of new and historic HF radar data. 15

Task 2. High-level quality control of new and historic HF radar data. 573 grid points available 16

Task 3. Mesoscale reanalysis and forecast of the atmospheric fields by Aladin/HR and WRF-ARW models. Which models to use for N-Ad study (reanalysis) and which for M-Ad study (operational, reanalysis)?  Aladin-HR vs. WRF/ECMWF vs. WRF/Aladin-SI vs. ... Intercomparisons of atmospheric models Case studies (bora, sirocco, ...) Milestone 4. Aladin/HR surface model fields available for the northern Adriatic area (month 6) and middle Adriatic area (month 14) covered by HF radar measurements.  done in month 4 (February 2014), thanks to Martina Milestone 5. WRF-ARW surface model fields available for the northern Adriatic area (month 8) and middle Adriatic area (month 16) covered by HF radar measurements.  done in months 5 and 10 (March and August 2014), thanks to Nedjeljka and Blaž

Task 3. Mesoscale reanalysis and forecast of the atmospheric fields by Aladin/HR and WRF-ARW models. http://meteo.fmf.uni-lj.si/NEURAL The case of bora in the lee of Biokovo Mountain, which was not reproduced by Aladin/HR but by WRF-ARW. http://meteo.hr 18

Task 4. Tuning the neural network parameters for optimal performance. Which SOM parameters will be used?  map size, lattice structure, neighbouring structure Parameters to be used by SOM analysis (surface winds), area covered? First to be done for N-Ad and later for M-Ad Milestone 6. Neural network parameters tested and the best parameters chosen, northern Adriatic site (month 9), middle Adriatic site (month 17).  done in month 9 (July 2014), thanks to Hrvoje K.

Task 5. Obtaining characteristic patterns of surface currents and meteo model data through learning process. Time series for N-Ad and M-Ad to be split into two equal series, one for training and another for assessment of forecasted values SOM solutions obtained for training series for both N-Ad and M-Ad Associating the SOM patterns to dominant winds and processes Comparison with traditional approach (conditional averaging, ...) Milestone 7. Characteristic patterns of surface ocean currents computed for both northern (month 10) and middle Adriatic HF radar sites (month 18).  done in month 11 (September 2014), thanks to Hrvoje K. and Hrvoje M. 20

Task 5. Obtaining characteristic patterns of surface currents and meteo model data through learning process. 21

Task 6. Assessing of the abilities and skill performance of the system. To create SOM-based hindcast for second part of the examined interval To compate hindcast solutions with the measured HF radar series To assess and quantify the quality of the forecast (rmse, complex correlation, ...)  evaluation of skill parameters  eventual recomputation of SOM solutions based on other SOM parameters Milestone 8. Skill parameters of the modelling systems computed and assessed (months 12, 20).  done in month 16 by Hrvoje K. and Hrvoje M.

Task 6. Assessing of the abilities and skill performance of the system. 23

Task 7. Setting up procedures and protocols for operational ocean forecast. To define operational procedures and flow-chart of the forecasting system To define protocols of the data flow from NWP model through choice of the best HF radar forecast solution up to its presentation on the web Milestone 9. Flow chart, procedures and protocols of the operational system developed (month 21).  done in month 23 by Damir.

Task 8. Creating dynamic web content with forecast of ocean currents. Project web pages (at the beginning of the project) How to manage the data within the database? How to visualise the forecasts? Dynamical programming and creation of the forecast web pages. Milestone 10. Dynamic web content with forecasts of ocean currents created (month 23).  done in month 24 by Damir.

http://faust.izor.hr/autodatapub/neural_som_karta?jezik=eng 26

Task 9. Dissemination of project results. Conferences, workshops International project/programme meetings (MonGOOS, EuroGOOS, European Marine Board, ...) Public (Festival of Science, public lectures, various media, ...) Stakeholders (national governmental authorities, operational services and civil protection agencies, ...)  stakeholder workshop to be organised at the end of the project Milestone 11. The information on project results and application possibilities will be systematically disseminated to potential users (months 12-24)  done by public (Festival of Science), science lectures (IRB, UniLj), intl. programmes (MonGOOS), conferences Milestone 12. Stakeholder workshop organized (month 24)  collocated with the Final Workshop on 29 September 2015 at the Institute.

Our obligations on the project: 1. To install and maintain HF radar system in the middle Adriatic - To be done in collaboration with HAZADR project  done in March 2014 2. To create dynamic web pages with ocean forecasts for N-Ad and M-Ad - That is the major outcome of the project  done in September 2015 3. To submit at least 3 WoS papers (but we can more) - Kalinić, H., Mihanović, H., Cosoli, S., Vilibić, I., 2015. Sensitivity of Self-Organizing Map surface current patterns to the use of radial versus Cartesian input vectors measured by high-frequency radars, Computers & Geosciences, 84, 29–36. Vilibić, I., Kalinić, H., Mihanović, H., Žagar, N., Jesenko, B., Tudor, M., Cosoli, S., 2016. Sensitivity of joint HF radar and ground wind Kohonen maps to various data processing procedures and mesoscale model inputs, Computational Geosciences, in press, doi: 10.1007/s10596-015-9550-3. Vilibić, I., Šepić, J., Mihanović, H., Kalinić, H., Cosoli, S., Janeković, I., Žagar, N., Jesenko, B., Tudor, M., Dadić, V., Ivanković, D., 2016. Self-Organizing Maps-based ocean currents forecasting system. Scientific Reports, almost accepted (minor revisions). Kalinić, H., Mihanović, H., Cosoli, S., Tudor, M., Vilibić, I.: Is it possible to predict ocean surface currents using numerical weather prediction model and Kohonen neural network? A northern Adriatic study, submitted to Appplied Soft Computing, December 2015

Our obligations on the project: 4. To submit at least 2 international projects and one national project related with NEURAL, of which one should include collaboration between PI and Co-PI institutions HRZZ project ODA  November 2013 H2020 BG9 project EMILIA  March 2014 ERC CoG project SINAI  May 2014 - HRZZ project SCOOL  September 2014 (approved!) - Interreg Central Europe project HARFA  June 2015 - H2020 project BG-10  operational oceanography in the Mediterranean - H2020 project BG-9  support to the BlueMed - Interreg ADRION (just opened)  at least one proposal - Interreg HR-IT (open in June 2016)  2-3 proposals ... 5. To organise stakeholder workshop Done on 29 September 2015 6. Dissemination (5 targeted presentations/public lectures) Lecture in IRB - Zagreb, Festival of Science - Split, MonGOOS annual meeting 2014 - Lecce, Croatian Meteorological Society Workshop 2014 - Zagreb, AGU2014 – San Francisco, University of Ljubljana

Our obligations on the project: Assessment by an individual consultant (expert) - First assessment in March 2015: Based on the recommendation, we compared the modelling performance and skill of the SOM-based forecast and “classical” ocean forecast by ROMS numerical model operational for the Adriatic Sea – basis for Scientific Reports paper - Final assessment (September 2015) 30

Key Performance Indicators: 1 1 1 1 1 1 3 2 3 6 9 1 2 1 1 31

Sugar at the end - alias finances: DIRECT COSTS planned spent Cons.+Dur.+Other 409.700,00 401.739,93 Personnel 560.000,00 580.939,80 Travel 152.150,00 139.169,93 Consulting 27.500,00 Overheads 202.826,00 202.826,35 Total 1.352.176,00 CO-FINANCING planned spent In-cash IOF 252.000,00 824.295,00 In-kind IOF 245.859,09 In-kind UniLj 199.320,00 211.490,00 In-kind MHS 150.000,00 - Total 601.320,00 1.281.644,09 32

Conclusions and perspectives The project reached a fair progress for a 2-year project (definitelly a nice news from the perspective of the PI):  technological: new and emerging technology (HF radars) installed and operational, high-resolution NWP system set up, lot of new data (meteo, ocean) collected and available for research  research: 3-4 articles are or will be (hopefully) published, more to come  operational: a brend new and innovative forecasting system is operational and on the web