700 Proceedngs of the 8th Internatonal Conference on Innovaton & Management Forecastng the Demand of Emergency Supples: Based on the CBR Theory and BP Neural Network Fu Deqang, Lu Yun, L Changbng School of economcs and Management, Chongqng Unversty of Post and Telecommuncaton, Chongqng, Chna, 400065 (E-mal: fudq@cqupt.edu.cn, luyun0622@sna.com, lcbss@126.com) Abstract: In recent years, the unconventonal emergence ncdents occurred frequently, whch are serously harmful to people s lves and property. So how to predct resource requrements after dsasters tmely becomes an mportant ssue. Ths paper presents an accurate predcton method based on CBR and BP neural network. Frstly, a set of base cases of the emergency demand should be found out. Then the law of emergency supples base cases can be obtaned accordng to BP neural network, at last, to prove the value of the forecast method, the demand of goal case can be predcted by usng the law. Key words: Emergency supples; Forecast; Case-based reasonng; BP neural network 1 Introducton Recently, the unconventonal emergence ncdent occurred frequently whch s serously harmful to people s lves and property. The researches of emergency supples demand forecastng play an mportant role n mprovng the emergency response effectveness and mnmzng the dsaster losses. Emergency supples means once emergence ncdent occurs, the smlar or alternatve resources whch can be moblzed. Currently, many scholars have conducted studes to forecast the demand of emergency supples. Guo Rupeng [1] combned the fuzzy reasonng wth the case-based reasonng and establshed the emergency supples demand model accordng to the characterstcs of the materal requrement. Ln Janxn [2], who establshed the mathematcal model to forecast the emergency transport demand dstrbuton problem, proposed an mproved emergency transport demand curve functon. Fu Zhyan [3] predcted the demand of emergency supples usng case-based reasonng model. Certan contrbutons have been made from ther studes. In concluson, now many scholars have done relevant researches of emergency supples demand forecastng from dfferent angles. The emergency events are characterstc of non-routne, sudden, and uncertanty; however, the prevous research dd not hghlght these characters, so these methods are of low practcalty. Ths paper presents an accurate predcton method based on CBR and BP neural network, and selects emergency supples samples whch are used n BP neural network tranng. Because smlar emergency supples demand processes are bound to have smlar demand law, frst, puttng forward a set of based cases whch are smlar wth the goal emergency supples demand, then fndng out the demand law of based cases by usng BP neural network, and fnally apply the law to predct the goal case of emergency demand. 2 Modelng 2.1 The framework of model The framework of emergency supples demand forecastng manly conssts of 3 modules: 1)Smlar case retreval module 2)Access to the law of demand module 3)Emergency supples demand forecastng module The functon of module 1) s to dentfy one case whch s smlar to the goal case n the case database; the functon of module 2)s to gan the projectve reconstructon between the characters of emergency supples and the demand; the functon of module 3)s apply the law from module2) to predct the demand of the goal case emergency supples by enterng the goal case character message nto the law. 2.2 Fundamentals Used n the Forecastng of Emergency Supples Demand 2.2.1 CBR (Case-based reasonng) Ths work s supported by key Laboratory of Electronc commerce and Modern Logstcs, Chongqng Educaton Commsson
Proceedngs of the 8th Internatonal Conference on Innovaton & Management 701 CBR s knd of methodology whch not only can mtate human reasonng and thought process, but also can buld ntellgent computer systems. It can smulate thought processes of experts, such as assocaton, ntuton, analogy, nducton, learnng and memory and so on, the core dea s that when solvng problem, prevous experence and acqured knowledge of smlar problems can be used to reason the new case, and adjust the soluton accordng to the dfferences of old and new, then the new soluton of the problem wll become a new case. The new case can be added nto the case database. As the growth of the case database, the experence of the system wll be more abundant. Ths paper consders the emergency supples demand as dependent varables. Frst of all, case database of emergency supples should be establshed. There are n cases, and suppose the number case s denoted by C, ( =1,2, n),namely, the base case set s C, C={C 1,C 2,,C n }. Therefore, assumng that there are m characters, and they are hghly assocated wth the demand of emergency supples. The character set s denoted by X, X={x 1,x 2,,x m }. Accordng to the dfferent needs of emergency resources, demand for emergency supples can be dvded nto 3 categores, manpower needs, materal needs, and fnancal needs. The demand of the goal case s denoted by D. Because smlar emergency supples demand always be based on the smlar demand law, n smlar cases can be chose as tranng sample. 2.3 BP Neural Network 2.3.1 Network structure BP neural network s a knd of error back propagaton learnng algorthm. It s composed of the nput layer, the hdden layer, output layer and the connectons between dfferent neurons, the layers of neurons s connected by weghts and thresholds. BP neural network s used to create the mappng between emergency supples demand and the character parameters, thereby emergency supples demand forecastng process can be solved by BP neural network. The process of emergency supples demand forecastng of BP neural network manly conssts of 3 parts: (1) Data collecton, enter the hstorcal demand data nto case database (2) Use emergency supples character data and the correspondng demand data to tran network. (3) Apply the network tranng results to forecastng the demand of emergency supples accordng to the character parameters. 2.3.2 Mathematcal Expresson BP neural network can be seen as a knd of nonlnear mappng from nput to output, that s: n m D: F = F, f( x) = Y, For a sample set: nput X R n, output Y R m, then there exsts a mappng. gx ( ) = y( = 1, 2,... n), Now map f s requested, so n some sense (usually the number of least squares), f s the best approxmaton of g. BP neural network s smlar to complex functon through several complex of smple functon. BP learnng algorthm s a very mportant and classcal learnng algorthm, whch can be used to acheve multlayer feed forward BP neural network weghts adjustment [4]. The gudng deology of the network learnng formula s that for the amendment of the network weghts (w j,t l ) and the threshold θ, the error functon E fall along the gradent drecton. 3-BP network node s expressed as: nput node x j, hdden node y, output node O l, the network weght of nput nodes and hdden nodes are w j, and hdden nodes and output nodes are T l. When the expected output of output nodes s T l, the BP model formula and ts dervaton s as follows: (1)The formula of output O l of output nodes Input of nput nodes: x j Output of hdden nodes: y = f( wj θ) j (1) Among them, the connecton weghts w j, the nodes thresholds θ, Output of output nodes: O = f( Tjy θ) (2)
702 Proceedngs of the 8th Internatonal Conference on Innovaton & Management Among them, the connecton weghts T j, the nodes thresholds θ l (2)The correcton formula of output layer (hdden nodes to output nodes) s the expected output of output nodes T l; All of the sample error: Among them, one sample error: n (3) k 1 k E = e < ε (4) n n ( k) ( k) k = k 1 1 1 1 e t o Among them, k s sample number; n s output nodes number. Error formula: Weghts correcton: δ = ( T O ) O( l O ) (5) l l l l Tl ( k+ l) = Tl ( k) + ηδl y (6) Whch, k s the teraton number; η s the learnng rate, η decdes the changes whch are produced n each tranng cycle. And t s generally rangng from 0.01 to 0.8; (3)Thresholds correcton: θl( k+ l) = θl( k) + ηδ l (7) The correcton formula of the hdden layer(nput nodes to hdden nodes) Error formula: Weghts correcton: Thresholds correcton: δ = y (1 y ) δ T (8) l l w ( k+ l) = w ( k) + η δ x (9) j j j θ ( k+ 1) = θ ( k) + ηδ (10) j Ths algorthm s an teratve process, each round wll adjust the weght agan untl the error of expected output and calculated output s less than an allowable value. Then end the learnng tranng, buld the model shown as Fgure 1. δ 1 δ l δ n Input layer T 1 T l T n Hdden layer Fgure 1 Back-propagaton Dagram 3 Examples 3.1 Index Select Takng the earthquake emergency supples demand forecastng for example, there are 30 cases n the case database shown as the Table 1, C,=(1,2,,30). Selectng the follow 4 chef characters of the earthquake emergency supples as the neurons of nput layer, X={x 1,x 2,x 3,x 4 }={Magntude, depth, densty of populaton, Number of affected people}, and output s the demand of earthquake emergency supples. Takng C(=1,2...,15) as learnng sample set, and use X as smulaton nput of network to
Proceedngs of the 8th Internatonal Conference on Innovaton & Management 703 learnng and tran. Takeng C(=16,17,...,30) as the test sample to examne the accuracy of neural network learnng, and compar the predcted values wth the actual values untl the errors are adjusted to a certan range, then forecast the demand of earthquake emergency supples usng the traned network. 3 categores emergency resources demand are all converted by yuan. Table 1 Case Database demand x 1 (rchter scale) x 2 (Km) x 3 (one person per Sq.Km) x 4 (10 thousand) (10 thousand yuan) C 1 6.0 10 76.11 20 77 C 2 6.1 18 81.43 13.7 3843 C 3 5.5 6 240.97 18 1970 C 4 5.0 6 166.7 41 318 C 5 5.0 6 48.18 3.9 1268 C 30 5.2 10 414 2.2 1209 Selectng sgmod functon as the transfer functon between the nput layer and the hdden layer,and lnear functon between the hdden layer and output layer. There s only 1 hdden layer, and the experental number of hdden nodes s 7. Fcture 2 s the dagram of the 3-layer BP neural network structure. x 1 x 2 x 3 x 4 O Input layer Hdden layer Output layer Fgure 2 3-layer BP Neural Network Structure Dagram 3.2 Tranng and Result Analyss Usng Matlab 7.0 to do BP neural network emulate tranng, and the max learnng rato s 0.3, the mnmum s 0.0. After 10000 teratve processes, the average error comes to the reasonable range, less than 5%. Fgure 3 s the network tranng parameter dagram. Fgure 3 Network Tranng Parameter SSW Dagram Under ths crcumstance, Suppose D, X D = (7.3, 8, 24, 100). Emergency supples demand of D can be predcted. The predcted value s 3429, and the actual value s 3290, the error s 4.2%. 4 Conclusons Accordng to above analyss, the method of Forecastng Emergency Supples Demand Based on the CBR Theory and BP Neural Network s feasble; the error between the predcted value and the actual value s under reasonable control. However, n order to make the forecastng more accuracy, the number of the case should be greater, and then a more well law can be obtaned. Reference [1] Guo Rupeng. Research on the Methods and of Emergency Materal Moblzaton Decson-makng [D]. Bejng: Bejng Insttute of Technology, 2006:34-40 (In Chnese) [2] Ln Janxn, We Xanlan, Wu Hayan. Tme-varaton Demand Forecast of Emergent Traffc Evacuaton Based on S-curve[J]. Journal of Transport Informaton and Safety,2009,27(3):92-96 (In Chnese) [3] Fu Zhyan, Chen Jan. Research on Emergency Materal Demand Forecast Model n Dsaster [J].
704 Proceedngs of the 8th Internatonal Conference on Innovaton & Management Logstcs Sc-Tech, 2009:11-13 (In Chnese) [4] Yn Chao, Qan Shengsan,Zeng Yonggang, ZhouYuan. Research for the Estmaton of Capacty Demand of Knowledge-producton Based on the CBR Theory and BP Neural Network [J]. Scence-Technology and Management, 2010:45-48 (In Chnese)