4 A publcaton of CHEMICAL ENGINEERING TRANSACTIONS VOL. 46, 05 Guest Edtors: Peyu Ren, Yancang L, Hupng Song Copyrght 05, AIDIC Servz S.r.l., ISBN 978-88-95608-37-; ISSN 83-96 The Italan Assocaton of Checal Engneerng Onlne at www.adc.t/cet DOI: 0.3303/CET54607 Tourst Arrvals Real-te Predcton Based on IOWA-Gauss Method Ln Chen, Maozhu Jn*, Yonghuan He Busness school, Schuan Unversty, Chengdu, 60000, Chna. jnaozhu@scu.edu.cn. Currently, the forecasts research focuses on tours n a tourst trends and toursts nfluencng factors. Although the forecast for the nter-annual and seasonal toursts has a wealth of research results, there s less study of everyday and real-te tourst arrvals. Ths paper analyzes the toursts real-te arrval law of Juzhagou, and then the use of herarchcal clusterng and Gaussan fttng atheatcal ethods for data processng, study the changes of day arrvals of tourst by segent analyss, and proposed a new odel for the predcton of toursts n real-te arrvals. We take Juzhagou Valley as an exaple to analyss, and experental results show that the forecast ethod s effectve.. Introducton The rapd developent of the tours ndustry has prooted local econoc developent, at the sae te, over-explotaton and the too large tours scale brngs enorous pressure to ecologcal and envronental protecton. Balance vstors scale, control of teporal and spatal dstrbuton of toursts has becoe an portant content of scenc area vstor anageent n peak travel perod. However, toursts dstrbuton are affected by season, te, weather, tourst type and other factors, so there s a bg uncertanty. There are varous ethods about forecastng tourst arrval, for exaple, Cho (003)fnd that the artfcal neural networks to forecast tourst arrvals perfor well n coparson to the exponental soothng ;Gl-Alana (005)forecast nternatonal onthly arrvals by usng seasonal unvarate long-eory processes ;Chu (008) uses fractonally ntegrated ARMA odels to forecast tours arrvals; S Chen (00) apply ANFIS odel to forecast the tourst arrvals to Tawan; H Song (0) forecast quarterly tourst arrvals by usng a new odel,the TVP-STSM. n addton, they also nclude ARIMA (Cho, 003; Cang, 0; Wan & Wang, 03), GARCH (Bollerslev,986; K & Wong, 006; Coshall, 009), SSA (Benek & Eeckels, 0), but we cannot fnd any paper adoptng a odel to forecast tourst arrvals of real-te dynac n the scenc area. Ths s need to study real-te dynac predcton of tourst nuber n the scenc area by Spatoteporal analyss of tours carryng capacty. So, the purpose of ths paper s to fll ths gap, ths paper cobne GAUSS algorth wth IOWA and buld a scenc area real-te predcton odel, segentatonly analyse the change of vstors on the scenc area, whch helps to prove the predcton ethod of the vstor nuber.. Model constructon Ths paper collects data through feld nvestgaton. Frstly process the data, fnd a general law of tourst arrvals. That s, the total nuber s gradually ncreasng, and the ncreasng speed s gradually slowng down, fnally the nuber tends to be stable. Next cluster data by usng herarchcal clusterng ethod accordng to sze and predct by usng the Gaussan fttng algorth. Then odfy predcton odel usng weght calculaton based on the proved IOWA operator. Obtan the fnal predcton results. Fnally, analyse the predcton results. Fgure s the research odel of ths paper. Please cte ths artcle as: Chen L., Jn M.Z., He Y.H., 05, Tourst arrvals real-te predcton based on owa-gauss ethod, Checal Engneerng Transactons, 46, 4-46 DOI:0.3303/CET54607
4 Fgure :The research odel 3. Eprcal study 3. Data preprocessng Observng the nuber of toursts who arrve scenc entrance at each te, we fnd t s a nonlnear nonstatonary te seres. However, f we ntegrally process the nuber of people of each oent,.e accuulate the nuber of people at each te, Let t denote te (nute), N denote the nuber of toursts who arrve t scenc entrance at each te, denote the total nuber of toursts arrvals, then: S = N, t =,,... t = t 3. Data clusterng accordng to sze The basc prncple of Herarchcal clusterng s: frstly, classfy a certan nuber of saples (or varables) nto ther own classes, then classfy two classes whch have the closest propertes nto a new class, calculate the dstance between the dfferent classes under the new class, cobne two classes whch have the closest propertes, repeat ths process untl all the saples (or varables) are cobned nto one class. Ths paper conduct herarchcal clusterng based on nuber scale, hopng that the scale dstance of data n a class s as close as possble, ake the average dstance of all tes n the cobned class s sallest.so we defne d( C, C ) = n Z Z j j Z C, Z j C j.that s Wthn-groups lnkage ethod. 3.3 Predcton odel based on Gaussan fttng algorth The days hstorcal data s dvded nto a nuber of scales. Ths paper fts the curve accordng to dfferent scales, gets a seres of predcton odels whch have sae structures, dfferent paraeters. In the actual fttng process, we do not requre y = f ( x) ( x strctly through all the pont, y), but requre the fttng error D = f( x) y n pont x s the sallest accordng to certan crtera. Often use the least squares approxaton searchng the best fttng curve ethod. y = F( x) Ths paper uses the gaussan functon as a basc functon of the curve. Naely, set as a gaussan functon syste, where each gaussan functon s deterned by three paraeters: peak heght A, peak poston B and peak wdth C. The entre gaussan functon syste s wrtten as: n x B F( x) = A exp = 3.4 Weght calculaton forecastng based on proved IOWA operator In.3, we calculated the nuber of people predcton odel under several scales. However, the sze s not entrely consstent wth the scale of the nuber of people arrval every day. When gven the total arrval nuber on a day, we hope to be able to predct the nuber of each oent arrval. Therefore, approprate weghts can be gven to the known szes to obtan the predcton value of each actual sze. The concrete steps are gven for solvng attrbute weghts accordng to ths thought as follows: j j Step: Calculate devaton dstance between dscrete scale and target scale j., v where represents the sze of scale. Step : Calculate the weght. C S t = j d j d = v v, =,,..., ω = dj, =,,..., d
ω Step 3: Sort order of the weght and the scale ( ω, ω,..., ω ), ω ω... ω d = dj, dj,..., dj, dj < dj <... < dj. f j = ω f, =,,..., f Step 4: Calculate the predcton equaton =, where represents the predcton equaton for scale. 4. Eprcal analyss In order to carry out perforance evaluaton on the proposed predcton odel, we conduct a nuber of eprcal studes on data based on real-lfe scenaros to predct the nuber of vstor arrval on each day and each te. 4. Data Sources Takng juzhagou scenc area as an exaple, we have collected data n real-te daly vstor arrvals fro May 0 to August 0(Collected fro RFID, n nutes).data collected fro 7:00 untl 3:00, altogether 70 nutes. Toursts scale s seasonal and toursts scale s slar when the date s close. In order to take all scales nto account as uch as possble n the forecastng process, t s napproprate to only let the prevous data as the tranng data. So randoly select several days as the tranng data, assess odel paraeters, and the rest s used to test the accuracy of the odel. 4. Scale herarchcal clusterng Conduct herarchcal clusterng on the nuber of days of tranng data, stratfed results are as the followng table (0 layers). Table : Herarchcal clusterng results layer 3 4 5 6 7 8 9 0 date 5.0 5.4 5. 5.3 5. 5.0 5.8 5.6 5.9 5.5 5.8 5.9 6.5 6.6 6.7 7.5 6.3 7. 6.9 6.30 5. 5.5 5.3 6.3 6.0 5.3 6. 8.9 scale 000 800 800 487 3600 0000 6300 5600 3 5000 layer 3 4 5 6 7 8 9 0 date 7.6 7.7 7.8 7.0 7.9 7.3 7. 7.6 7.7. 7.3 7.7 8.6 = > > > ( ) 7.4 7.5 7.0 7.8 7.30 7.3 7.9 8.4 8.8 8.3 8.7 8.8 8.4 8.7 8.5 8.0 8. scale 9000 000 000 5300 3500 900 400 6300 7700 8500 The scale of foraton classfcaton s arthetc ean of the all day data n a class, arrval nuber at each te under the new scale s also arthetc ean of the data at each te, changng fro 0 4 to.9 0 4 spans n scale. The lne graph of dfferent szes s as the followng fgure. The abscssa represents te (n nutes), the ordnate represents the total nuber of the arrval at each te. Fro lower to upper curve represents scale s ncreasng n turn. We fnd there s a followng regularty: T 43 Fgure : Vstor arrval curve The graph curve of dfferent szes exst a hgh degree of slarty and consstent trend. And ths curve shows that the growth rate of the nuber of vstor arrvals ncreases along wth te ncreased at the begnnng. After a growth perod, value of the growth rate gradually slow down and close to a certan lt. 4.3 Segented gaussan fttng Of coon fttng equaton, gaussan fttng have the hghest fttng accuracy to ths study. Therefore, the followng data of dfferent szes were gaussan ftted respectvely. After testng, 8 gaussan fttng errors s the
44 sallest n the gaussan fttng process. F( x)= So we choose eght forula gaussan fttng as fttng functon. Equaton s expressed as follows: A* exp( (( x B Descrbe t n detal by usng.8 0 5 )/ C) ) + A* exp( (( x B)/ C) ) scale as an exaple. Fgure 4 are the fttng curve and the resdual curve and the resdual s large, so we consder fragentng to reduce the fttng resduals. Fgure 3: The fttng and resdual curve Fgure 4: The segented curve In order to ensure the reasonableness of segentaton, we conduct frst, second dervatve on tourst arrvals growth curve. Shown n Fgure Introduce tourst destnaton toursts growth "speed" and "acceleraton" concepts and ts related varables. Dvde t nto three stages, and the segentaton ponts are 0 and 80. Use the Matlab software to ft, we can obtan ft equaton. The paraeters of the three segents of gaussan fts are as follows: Table : The three segents paraeters of gauss equaton Coeffcents (wth 95% confdence bounds) The frst segent (0-0) The second segent (0-80) The thrd segent (80-70) A 667.3 55.4.7E+04 B 4.3 8.3 787. C 0.04 3.96 056 A -9.35-7.944 0.7 B 05 63. 7.9 C 0.6545 5.696 7.678.. A8 89.6.9E+04 65.3 B8 05.3 90.8 374.5 C8 3.899 8.77 45.4 The frst segent Goodness of ft: SSE:.5e+004; R-square: 0.9999; Adjusted R-square: 0.9999; RMSE: 6.4. The second segent Goodness of ft: SSE: 45; R-square: ; Adjusted R-square: ; RMSE: 0.3. The thrd segent Goodness of ft: SSE: 699; R-square: 0.9998; Adjusted R-square: 0.9998; RMSE:.9. It shows that Gaussan fttng can pass the conforance testng and have hgh fttng precson to ths study. Repeat the above experent,0 paraeters values of dscrete scale equaton can be obtaned. Here s not to lst. 4.4 Calculaton of weghts The closer the scale s, the closer the curve graph s. To reduce the nterference by the too large dstance of the scales to the scale to be predcted, we conduct a secondary clusterng, and cluster dfferent scales nto four types. The result s as the followng table. Table 3: Secondary clusterng results Classfcaton Scale 0000 000 800 800 3600 487 5000 5600 6300 3 9000 000 3 000 900 3500 400 4 5300 6300 7700 8500 Ths paper choose data fro May 0, June 3, July, August 8 to predct the above four classfcatons respectvely to test the valdty of odel predcton.
45 fro the hstorcal data can be known, the tourst nuber on May 0, June 3, July and August 8 are respectvely.044 04,.430 04,.6 04 and.688 04 (unt: person) the four-day Predctng scales are respectvely recorded as F F F F, accordng to proved IOWA operator, we can get: 3 4 F=0.006 f+0.6554 f+0.34359f3 () ( f f and f 3 respectvely represent the gaussan fttng functon forulas of.0000 04,.000 04 and.800 04 n scale) F=0.000 f4+0.364 f5+0.6649 f6+0.000 f7+0.364 f8+0.000f9 () F3=0.00077 f0+0.0393f+ 0.85 f+0.47983 f3+0. f4+0.0970 f5+0.094 f6 (3) (4) F =0.5356 f +0.68 f +0.44 f +0.0000f 4 7 8 9 0 4.5 Analyss of experental results The followng s the experental results, the blue lne s the accuulated value each te of the actual data, the green lne s the predcton data. (a), (b), (c), (d) n the fgure 5 respectvely represent the real and predcted data on May 0, June 3, July, August 8;(a), (b), (c), (d) n the fgure 6 respectvely represent every nute error curve on May 0, June 3, July, August 8. (a) (b) (c) (d) Fgure 5: Real data and predcted data (a) (d) (c) (d) Fgure 6: Every nute error curve Fro Fgure 6 (a), (b), (d) can be seen: every nute errors on May 0, July and August 8 are respectvely n 0 or less; n the (c),despte three errors per nute value on June 3 are other than 0, but the overall trend s n 0 or less. It shows the valdty and relablty of the odel our proposed. 5. Concluson and prospect Ths paper proposes a new research queston about the toursts arrval predcton, and enuncates the necessty and portance of ths research queston fro a ore detaled te scale..e studes the nuber of toursts who reach the scenc area at each te per day. We fndthe graph curve of dfferent szes exsts a hgh degree of slarty and consstent trend and ths curve shows that the growth rate of the nuber of
46 vstors arrval ncrease along wth te ncreent at the begnnng. After a growth perod, value of the growth rate gradually slow down and close to a certan lt. In order to study ths proble deeply, on the base of stage predcton theoretcal odel based on tours theory and clusterng theory, ths paper buld a phased Gaussan fttng and weghts cobned forecastng odel to predct toursts arrval, and analyse and test the odel usng Juzhagou scenc area as an exaple. Exaple verfcaton shows that the proposed new odel has a good predctve accuracy. However, we do not take the errors after fttng nto account n the research process. The error can be used to correct n the future to prove accuracy. Then due that we solve a new proble to forecast real-te tourst nuber n the scenc area. The proposed predcton odel for ths new research queston does not conduct copared predcton wth other odel n tourst arrvals fled. Ths predcton odel and predcton ethod wll be proved n future studes. Acknowledgeents Ths work was supported by the Major Internatonal Jont Research Progra of the Natonal Natural Scence Foundaton of Chna (Grant no. 7000707), and the Natonal Natural Scence Foundaton of Chna (Grant no. 700075) and Huantes and Socal Scences project of The Mnstry of educaton of Chna (Grant no.yjc63003). References Bollerslev, T., 986. Generalzed autoregressve condtonal heteroskedastcty. Journal of Econoetrcs, 3(3), 307-37. DOI: 0.06/0304-4076(86)90063- Broohead, D. S., & Kng, G. P., 986, Extractng qualtatve dynacs fro experental data. Physca D: Nonlnear Phenoena, 0(-3), 7-36. DOI: 0.06/067-789(86)9003-X Cang, S., 0. A non-lnear tours deand forecast cobnaton odel. Tours Econocs, 7(), 5e0. DOI: 0.5367/te.0.003 Chen M. S., Yng L. C., Pan M. C., 00, Forecastng tourst arrvals by usng the adaptve network-based fuzzy nference syste [J]. Expert Systes wth Applcatons, 37(): 85-9. DOI: 0.06/j.eswa.009.06.03 Chu, F. L., 008, A fractonally ntegrated autoregressve ovng average approach to forecastng tours deand. Tours Manageent, 9(), 79-88. DOI: 0.06/j.touran.007.04.003 Coshall, J. T., 009, Cobnng volatlty and soothng forecasts of UK deand for nternatonal tours. Tours Manageent, 30(4), 495-5. DOI: 0.06/j.touran.008.0.00 De Goojer, J., Hyndan, R., 006. 5 years of te seres forecastng. Internatonal Journal of Forecastng, (3), 443 473. DOI: 0.06/j.jforecast.006.0.00 Gl-Alana L. A., 005, Modellng nternatonal onthly arrvals usng seasonal unvarate long-eory processes[j]. Tours Manageent, 6(6): 867-878. DOI: 0.06/j.touran.004.05.003 K, S. S., & Wong, K. K., 006, Effects of news shock on nbound tourst deand volatlty n Korea. Journal of Travel Research, 44(4), 457-466. DOI: 0.77/0047875058946 Song H., L G., 008, Tours deand odellng and forecastng A revew of recent research [J]. Tours Manageent, 9(): 03-0. DOI: 0.06/j.touran.007.07.06 Wan S. K., Wang S. H., Woo C. K., 03, Aggregate vs. dsaggregate forecast: case of Hong Kong. Annals of Tours Research, 4, 434-438. DOI: 0.06/j.annals.03.03.00