169_175 adg v–5 n01 Stanisl.pdf ANNALS OF GEOPHYSICS, VOL. 45, N. 1, February 2002 169 Forecasting of ionospheric characteristics during quiet and disturbed conditions Iwona Stanisl/awska and Zbigniew Zbyszyński Space Research Centre, Polish Academy of Sciences, Warsaw, Poland Abstract An autocovariance forecasting procedure for single location ionospheric characteristics is presented. Its accuracy is illustrated as a function of the amount of time extrapolation for selected European stations under quiet and disturbed conditions. 1. Introduction Day-to-day and hour-to-hour ionospheric variations are generally irregular. Irregular ionospheric variations are variations that cannot be predicted by any linear prediction method. They are caused by irregular changes in am- plitudes of ionospheric parameters as well as their time spread. In order to predict them much interest and efforts have been dedicated in the past. Various non-linear techniques have been proposed. The latest publications deal with a statistical approach (Muhtaov and Kutiev, 1999), or with a modern neural network technique (Wintoft and Cander, 1999, 2000; Tulunay et al., 2000). This paper presents a continuation of this work (Stanisl/awska and Zbyszyński, 2001) in the application of the autocovariance prediction method for ionospheric purposes. 2. The autocovariance prediction method application The autocovariance prediction method was originally elaborated at the Space Research Centre for prediction of irregular variations in Earth rotation (Kosek, 1993, 1997). In this method the first prediction point outside the data time interval in the future and in the past is computed and added at the beginning or at the end of data, respectively, so the next prediction point can be computed. The difference between the prediction and data at a particular time in the future computed at different starting prediction epochs reveals unpredictable or irregular variations of the considered ionospheric parameter. Along this line the analysis of potential application of this method in the ionosphere was presented in Stanisl/awska (1994). The forecasting capabilities of the method for f0 F2 parameter were shown in Stanisl/awska and Zbyszyński (2001). This paper presents a similar analysis for other ionospheric characteristics as well as the forecast dependence on the time range considered. One of the advantages of the autocovariance method is that it does not require any addition- al parameters, which describe solar and/or geophysical conditions. It also means that any Mailing address: Dr. Iwona Stanisl/awska, Space Re- search Centre PAS, 00-716 Warsaw, Bartycka 18a, Poland; e-mail: stanis@cbk.waw.pl Key words forecasting of ionospheric characteristics 170 Iwona Stanisl/awska and Zbigniew Zbyszyński additional uncertainty connected with the need to use a prediction of these is avoided. The only information needed is a long enough period of observation. Another advantage is that it can be used for forecasting without any knowledge of the morphological and physical processes in a medium, such as the ionosphere. 3. Results and conclusions The present study investigates the f0 F2, f 0F1, f 0 E, and M(3000)F 2 parameters. Data have been taken from RAL-CD-ROM, prepared by the Rutherford Appleton Laboratory, United Kingdom, for European Cooperation in the Field of Scientific and Technical Research (COST) Action 251 (Hanbaba, 1999) and from the Ionospheric Despatch Centre in Europe (Stanisl/awska et al., 1999) (http://www.cbk.waw.pl/rwc/idce.html), that is a COST 251 initiative. A list of relevant stations is presented in table I. In this method the prediction estimation is computed as a function of an observed variable (different ionospheric characteristic) from several ionospheric stations only. The sampling interval of ionospheric characteristics, in our case, is 1 h. In this paper, 1-, 2-, 4-, 8-, 12-, 24- and 48-h-ahead forecast have been obtained. Particular purposes of this application are to consider the needs of the forecast obtaining for instantaneous situation by operational service of the ionospheric situation. To deal with a real situation any specific method for gaps filling has been used. Any data gaps were replaced by 7-days-smoothed average only. Also the requirements of the time period of available data were limited. Data from separate periods within September - November 1998 were used. The number of data used in the computation for the presented statistics is shown in table II. For the calculations of the short time forecast (up to 12 h ahead) 73 past values (3 days) hour by hour, have been taken, while for a longer time forecast (24-48 h ahead) 28 values from the previous 28 days, for each hour separately, have been taken. To satisfy the requirements of the method, as input data we used deviations of the measurements from the 7-days-smoothed Table I. Stations and their geographical coordinates. Location Station Latitude, °N Longitude, °E Tortosa EB040 40.8 0.5 Rome RO041 41.9 12.5 Juliusruh JR055 54.6 13.4 Sofia SQ143 42.7 23.4 Warsaw MZ152 52.2 21.2 Uppsala UP158 59.8 17.6 Lycksele LY164 64.6 18.8 Kiruna KI167 67.8 20.4 Table II. Number of considered data points. Disturbed Quiet Total f0 E 1900 3900 5800 f 0 F 1 500 900 1400 f 0 F 2 6100 9200 15300 M(3000)F2 6100 8500 14600 171 Forecasting of ionospheric characteristics during quiet and disturbed conditions Fig. 1. Bar chart created for RMS errors for f0 F2 characteristic, for disturbed data (upper panel), quiet data (middle panel), and for all data together (lower panel). Autocovariance forecasting method: For1, 1 h ahead For2, 2 h ahead, etc. Persistence: P1. Fig. 2. The same as fig. 1 for f0 F1 characteristic. Fig. 3. The same as fig. 1 for f0 E characteristic. Fig. 4. The same as fig. 1 for M(3000)F 2 characteristic. foF2 Dist urbed Dat a F o r1 F o r2 F o r4 F o r8 F o r1 2 F o r2 4 F o r4 8 P 1 IT U -R 0.2 0.8 1.4 foF2 Quiet Dat a F o r1 F o r2 F o r4 F o r8 F o r1 2 F o r2 4 F o r4 8 P 1 IT U -R 0.2 0.8 1.4 foF2 All Dat a F o r1 F o r2 F o r4 F o r8 F o r1 2 F o r2 4 F o r4 8 P 1 IT U -R 0.2 0.8 1.4 foF1 Dist urbed Dat a F o r1 F o r2 F o r4 F o r8 F o r1 2 F o r2 4 F o r4 8 P 1 IT U -R 0.1 0.3 0.5 foF1 Quiet Dat a F o r1 F o r2 F o r4 F o r8 F o r1 2 F o r2 4 F o r4 8 P 1 IT U -R 0.1 0.3 0.5 foF1 All Dat a F o r1 F o r2 F o r4 F o r8 F o r1 2 F o r2 4 F o r4 8 P 1 IT U -R 0.1 0.3 0.5 foE Dist urbed Dat a F o r1 F o r2 F o r4 F o r8 F o r1 2 F o r2 4 F o r4 8 P 1 IT U -R 0.0 0.2 0.4 foE Quiet Dat a F o r1 F o r2 F o r4 F o r8 F o r1 2 F o r2 4 F o r4 8 P 1 IT U -R 0.0 0.2 0.4 foE All Dat a F o r1 F o r2 F o r4 F o r8 F o r1 2 F o r2 4 F o r4 8 P 1 IT U -R 0.0 0.2 0.4 M(3000)F2 Dist urbed Dat a F o r1 F o r2 F o r4 F o r8 F o r1 2 F o r2 4 F o r4 8 P 1 IT U -R 0.05 0.15 0.25 M(3000)F2 Quiet Dat a F o r1 F o r2 F o r4 F o r8 F o r1 2 F o r2 4 F o r4 8 P 1 IT U -R 0.05 0.15 0.25 M(3000)F2 All Dat a F o r1 F o r2 F o r4 F o r8 F o r1 2 F o r2 4 F o r4 8 P 1 IT U -R 0.05 0.15 0.25 1 2 3 4 172 Iwona Stanisl/awska and Zbigniew Zbyszyński average value calculated for each hour sep- arately. An additional correction factor for 1-h- ahead forecast is introduced in the following manner: for the forecast at time t, parameters at times t-1 and t-2 are calculated. If the values in t-1 and t-2 differ from the measurements by more than 30%, the current calculated forecast is changed exactly by the last value of the gradient. The forecast was analysed for the disturbed and quiet periods separately. For the distinction between quiet and disturbed conditions, the catalogue of disturbances (Kouris et al., 1998) observed at selected ionospheric stations was used. The catalogue is available at the Ionospheric Despatch Centre in Europe. The prediction error is presented as an RMS error and percentage deviation of the prediction against the measurements. For comparison with the forecast results using the ITU-R-predicted median value and the measurement from the previous hours (persistence) have also been given. Figures 1 to 4 show the bar charts of the RMS error for four considered characteristics for quiet and disturbed conditions separately, as well as for all data together. The method has been used Fig. 5. 1- and 24-h-ahead forecast, and measurements for f0 F2 characteristic for three 5-day-periods in September 1998 at Kiruna station. 173 Forecasting of ionospheric characteristics during quiet and disturbed conditions Fig. 6. The same as fig. 5 for f 0 E characteristic at Uppsala station. for 1-, 2-h ahead, etc. Samples of 1-, and 24-h- ahead forecast for f0 F2, f0 E and M(3000)F2 characteristics for the most quiet, incidentally chosen moderate disturbed and the most disturbed periods are shown in figs. 5 to 7. Five day periods in October and September 1998 observations, 1- and 24-h ahead, are presented. Incidentally chosen, but representative for the numerical experiments performed, stations from mid- and high latitudes are presented. Only the plot for f0 F1 characteristic (fig. 8) is presented for a 5-day-period in September 1998 for a mixed quiet and negatively disturbed period for the mid-latitude Warsaw station only. For all the considered data the forecast shows much better results than medians and per- sistence. Except for the most disturbed data for f 0 F 2 characteristic, when the persistence is better than the forecast 2-, 4-, etc. hours ahead. But obviously, while considering the forecast for longer than 1-h-ahead period, the persistence does not exist. It ought to be mentioned that ITU-R prediction is never available with such accuracy, as used in this paper, because of the actual solar activity parameters used (not prediction). For M(3000)F 2 characteristic, increasing the time range of the forecast also increases the errors. This increase is not so 174 Iwona Stanisl/awska and Zbigniew Zbyszyński Fig. 7. The same as fig. 5 for M(3000)F2 characteristic for three 5-day-periods in October 1998 at Juliusruh station. Fig. 8. The same as fig. 5 for f0F1 characteristic for one 5-day-period in September 1998 at Warsaw station. 175 Forecasting of ionospheric characteristics during quiet and disturbed conditions pronounced as for f 0 F 2 . For f 0 F 1 and f 0 E characteristics the situation is quite different. The impact of the data from sunrise and sunset hours, particularly for f 0 F 1 data, and night hours for f 0 E, enlarge the errors, because the forecast is sometimes given for non-existing F1 layer, and vice versa. Generally, the quiet situation, which is represented by much more smoothed data than for the disturbances, is predicted with higher accuracy. During the disturbance lasting several hours, the method might generate some rapid fluctuations. When the method gives too high, or too low values, the correction by gradients improves the forecast, but only for the second, and higher measured values hours. So the forecast curve for later hours is much smoother and closer to observations. However, for individual disturbances the errors might still be substantial. This effect can be avoided using data with higher resolution sampling, as 15, or 5 min. In such a case 1-h-ahead forecast will follow the observations with higher accuracy after half an hour, or 10 min, respectively. Generally, the autocovariance method shows the correctness of this approach for any ionospheric characteristics. The autocovari- ance method of ionospheric characteristics forecasting provides an acceptable accuracy. This might also be the crucial point for predicting the electron concentration height profiles. Its quite reliable results for quiet, as well as for disturbed conditions allow us to conclude that this method can be used in operational services of ionospheric situation, as used to update the limited-area ionospheric forecast at the Regional Warning Centre Warsaw of the International Space Environment Service. Acknowledgements This research was partly supported by Polish Committee of Scientific Research grant 2 P03C 006 17. REFERENCES HANBABA, R. (1999): COST 251 Final Report, SRC Print. Off., Warsaw. KOSEK, W. (1993): The autocovariance prediction of the Earth rotation parameters, in 7th International Symposium «Geodesy and Physics of the Earth», IAG Symposium No. 112, Potsdam, Germany, October 5-10, 1992, edited by H. MONTAG and CH. REIGBER (Springer Verlag), 443-446. KOSEK, W. (1997): Autocovariance prediction method of short period Earth rotation parameters, Artif. Earth Satell., 32 (2), 75-85. KOURIS, S.S., D.N. FOTIADIS and T.D. XENOS (1998): On the day-to-day variation on f 0 F 2 and M(3000)F 2 , Adv. Space Res., 22 (6), 873-876. MUHTAROV, P. and I. KUTIEV (1999): Autocorrelation method for temporal interpolation and short-term prediction of ionospheric data, Radio Sci., 34 (2), 459-464. STANISL/ AWSKA, I. (1994): A single-station prediction model as a contribution to instantaneous mapping, Ann. Geofis., 37 (2), 153-157. STANISL/ AWSKA, I. and Z. ZBYSZYŃSKI (2001): Forecasting of the ionospheric quiet and disturbed f 0 F 2 values at single location, Radio Sci., 36 (5), 1065-1071. STANISL/ AWSKA, I., T.L. GULYAEVA and R. HANBABA (1999): Ionospheric Despatch Centre in Europe, Phys. Chem. Earth (C), 24 (4), 355-357. TULUNAY, E., C. OZKAPTAN and Y. TULUNAY (2000): Temporal and spatial forecasting of the f0 F2 values up to twenty four hours in advance, Phys. Chem. Earth (C), 25 (4), 281-285. WINTOFT, P. and LJ.R. CANDER (1999): Short-term prediction of f0 F2 using time-delay neural network, Phys. Chem. Earth (C), 24 (4), 342-347. WINTOFT, P. and LJ.R. CANDER (2000): Ionospheric f 0 F 2 storm forecasting using neural networks, Phys. Chem. Earth (C), 25 (4), 267-273.