TW200404195A - Prediction system - Google Patents

Abstract

A prediction system comprises a prediction model forming unit 1 for predicting the subject of prediction, including a prediction model having a plurality of parameters; a parameter value storage unit 2 for storing a plurality of parameter values; and an input-and-output data storing means for storing time-series input and output data. A static parameter identifying means determines a restrictive condition to be imposed on parameters, on the basis of stationary input and output data. A dynamic parameter identifying means determines the values of the parameters within a range meeting the restrictive condition determined by the static parameter identifying means on the basis of nonstationary input and output data stored in the input-and-output data storage means, and sends the values of the parameters to the parameter value storage means.

Description

200404195 (1) 发明 Description of the invention [Technical field to which the invention belongs] The present invention relates to a prediction system, which includes a prediction model for online prediction of natural phenomena and man-made phenomena. For example, the prediction of the rate of rainwater flowing into a sewage plant, Forecast of inflow quality, forecast of river level, forecast of dam storage capacity, and forecast of demand for water, natural gas and electricity. [Previous technology] Forecasting technology is one of the important basic technologies mainly used in the field of utilities. It is used to predict the demand for water and the flow rate of rainwater into sewage plants. Predictive control, and economic sectors used to predict stock prices. In general, forecasts are obtained by processing the data available at the start of some processing (hereafter referred to as "current time"). Typical prediction methods use some prediction models to make predictions. Prediction methods are roughly divided into the use of prediction models including numerical equations (white box approximation) based on laws of physics and chemistry, and the use of statistical processing based on real data, such as neural network learning, etc. , Or a prediction model constructed by system identification (black box approximation). The advantage of the former prediction method is that because of using the laws of physics (or chemistry), it is possible to make predictions that conform to (not conflict with) the laws of physics (or chemistry), and because in most cases the parameters have physical and chemical meanings So, you can adjust the model parameters quite easily. On the other hand, the disadvantage of the previous prediction method is that it is difficult to determine a parameter determination guideline when there are many parameters to be determined, that is, a priority guideline for deciding a parameter for adjustment. As a result, accurate predictions cannot be made. Another disadvantage of the previous prediction method is that the variables representing the relationship between cause and effect are neatly divided into cause variables (inputs) and result variables (outputs), and because in most cases no previous information about the result variables is used Therefore, the prediction accuracy will be reduced. The latter prediction method is to construct a model based on a predetermined deductive method or identification deductive method so that the model conforms to real data. Therefore, the model is at least suitable for accurate identification or learning of data, and can make accurate predictions even when the data is not used for most of the learning or identification. On the other hand, the disadvantage of the latter method is that when it is desired to change (adjust) the parameters to improve the prediction accuracy of the specific input data corresponding to the specific input data, it is impossible to adjust or fine-tune the decision through learning or recognition based on a predetermined deductive method. Parameters, because the parameters do not have physical and chemical significance and the guidelines for adjusting the parameters are unknown, readjustment or fine-tuning is not possible, and the predicted results will contradict the physical (or chemical) laws that must be met in accordance with the laws of nature. law. Prediction that violates the laws of physics such as conservation of mass or conservation of energy is one of the factors that predicts the black box approximation. The conventional prediction technology recognizes the advantages and disadvantages of the white box approximation method and the black box approximation method, and selectively uses the white box approximation method and the black box approximation method depending on the situation. The inventor of the present invention proposed a solution in the Japanese Patent Application No. 1 3 2998/200 1 "Apparatus for Adjusting Parameters of Process Model, and Apparatus and Method of Assisting Adjustment" (3) (3) 200404195 One of the shortcomings of the white box approximation is the problem caused by inadequate guidelines for parameter adjustment. This previously proposed method of parameter adjustment is based on the analysis of the sensitivity of the parameters to adjust the parameters. The unsatisfactory prediction accuracy due to unused past information related to the outcome variables can be improved by correcting the forecast 根据 based on past data related to the outcome variables. For example, when describing the prediction mode in a differential equation: .......... (1) y = 1ι {χ, θ, ΐί) .......... (2) Borrow By setting an appropriate observer, the prediction 値 can be corrected, and by using past information about the result variable (y), predictions can be made. In expressions (1) and (2), X, θ5 u, y, f, and h are the state variables of the model, parameters, input (cause variable), output (result variable), (non-linear) vector, and (Non-linear) output function. When constructing the pre-model by white-box approximation, we have tried to overcome the shortcomings of the white-box approximation by the methods described above. Many situations require the prediction model to be constructed by the black-box approximation. In most cases, it is difficult to apply the white box approximation to the specific forecast of the (forecast) system. In the system, for example, the outcome variables that need to be predicted, such as stage, flow, and water volume, and the reason variables such as rainfall are To some extent, it is unsightly, but internal structures such as the outflow treatment of rainwater are quite unsightly. In this case, it is preferable to construct a prediction model by a black box approximation. In particular, when considering a large amount of time series data stored in a data server or the like, in order to effectively use the data, the prediction using a black box model is more certain. From this point of view, the inventors of the present invention in Japanese Patent Application (4) (4) 200404195 please apply for 61866/1990 "Apparatus for Prediction Amount of Inflowing Rainwater, and Method of Predicting Amount of Inflowing Rainwater", Japanese Patent Application No. 224 766/1998 "Apparatus for Prediction Amount of Inflowing Rainwater, and Method of Predicting Amount of Inflowing Rainwater", and Japanese Patent Application No. 2423 1/2002 "Apparatus for Predicting River Stage" proposed some prediction methods using a black box model. These prediction methods have feature points and can actually provide highly accurate predictions. However, these methods are not free from the above disadvantages of the black box approximation. Therefore, when these methods cannot obtain accurate predictions or when they need to improve the pre-measurement accuracy of specific output data related to specific input data, the guidelines for reading and adjusting the identified parameters are not '凊It is possible to predict that the physical (chemical) laws represented by the conservation law are not met. [Summary of the Invention] The invention has been made in consideration of the above circumstances, and therefore, the object of the present invention is #MTypical i: a system using a discrete-time system model for the black box approximation method _ W force method to provide a prediction system, which It is possible to construct a prediction model that considers the laws of physics (the law of conservation) with a white-box approximation. Feng Gen According to the present invention, a prediction system for predicting a predictive target includes: a predictive model forming unit for predicting a predictive target, including a model with multiple parameters t ® 1; a parameter 値 storage unit for storing multiple Parameter 値; Input Λ & _ til Data storage mechanism for storing time series input (5) (5) 200404195 input and output data for prediction; static parameter identification mechanism, assuming static response (output) to fixed input Or, based on the static output data corresponding to the fixed input data, determine the limiting conditions to be added to the parameters; and the dynamic parameter identification mechanism, based on the non-static input data and output data stored in the input and output data storage mechanism, Determine the parameter 符合 within the range that meets the limit conditions determined by the static parameter identification mechanism, and transmit the parameter 给 to the parameter 値 storage mechanism. Unlike the conventional method of directly identifying dynamic parameters, the present invention uses a static parameter identification mechanism that takes into account the physical constraints of the steady state, puts constraints on the parameters, and can predict so that the steady state prediction is accurate The degree will not decrease. Since the relationship between the steady-state input and output corresponds to the law of physical conservation or in most cases in the prediction system, so it can constitute a prediction model that considers the laws of physics indirectly. Therefore, even in an unstable state, it is possible to avoid making predictions that conflict with the law of conservation, and to ensure that a physically acceptable prediction is obtained. In the prediction system according to the present invention, the prediction model forming unit includes a coefficient parameter section (automatic regression section) of a prediction model that works linearly on time series output data and a coefficient parameter section (moving average) that works linearly on time series input data And the static parameter identification mechanism, which applies constraints to the parameter 値 so that the ratio bb / aa is fixed, where aa is the sum of the coefficients 自动 of the automatic regression section, and bb is the coefficient of the moving average section 値In total. According to the present invention, the necessary conditions to be satisfied by a prediction model that conforms to the conservation of quality can be expressed by a very simple -9-(6) (6) 200404195 single exponent (restriction condition) by using a discrete-time linear transfer function model. This limitation The ratio between the sum of conditional parameters 値 is fixed. The discrete-time linear transfer function model includes the ARX model, ARMAX model, and BJ model, and is mainly used as a prediction model using the ink tank model. Therefore, by adjusting the parameters to meet the constraint condition that the ratio between the sums of the parameters 固定 is fixed, a black box approximation method can be used to construct a physical prediction. When needed, it is easy to change one of the adjusted parameters of the prediction model based on a simple index. In the prediction system according to the present invention, the static parameter recognition mechanism imposes a restriction condition such that the total number aa of the parameter 値 of the automatic regression unit is a fixed positive 値 which is very close to zero, and is expressed as aa = ε, 0 < ε < < 1. According to the present invention, the parameters can be adjusted so that the sum of the parameters 値 of the automatic regression section and the sum of the parameters 移动 of the moving average section remain fixed, and very promising adjustments can be achieved. In addition, predictions can be made so that the accuracy of predictions can be limited to the smallest degree of unavoidability even when a fixed or very slowly changing substance acts on the prediction system. Satisfactory prediction results can be given to a biased prediction system so that even if the input to the prediction system is zero, the output will not fall to zero. In the prediction system according to the present invention, the dynamic parameter determination mechanism recognizes the parameter 根据 according to the limiting condition, which requires the parameters 値 of the automatic regression section and the moving average section previously determined by the static parameter identification mechanism to be used respectively by using The selected best mechanism is fixed using non-stationary input and output data stored in the input and output data storage mechanism. According to the present invention, first determine the limiting conditions that should be met in any situation 'instead of directly identifying the dynamic parameters used by the conventional black box approximation. -10- (7) (7) 200404195, the limiting conditions are stored and the parameters Will be identified to match input and output dynamic data. Therefore, it is possible to prevent the accuracy of recognition from being lowered. Result 'Even under the condition that the general identification method will cause the number of parameters to diverge, the identification can still be obtained to improve the robustness of the prediction. For example, when the conventional adaptive identification method represented by sequential least squares is used to sequentially improve the accuracy of prediction, the parameters to be identified may be divergent and if the identification data contains abnormal data, no order can be made. Satisfactory forecast. When the method according to the present invention is used for adaptive prediction, since a limitation is imposed on the parameter 値 total, this phenomenon can be suppressed, and thus a prediction device having strong tolerance to abnormal data can be constructed. Although sufficient data cannot be obtained in most cases when predicting natural phenomena such as river water levels and sewage flow rates, these situations can still constitute a certain degree of reliable prediction device. According to the present invention, a prediction system for predicting a prediction target includes: a prediction model forming unit for predicting the prediction target, including a prediction model having a plurality of parameters; a parameter 値 storage unit for storing a plurality of parameters 値, an input and The output data storage mechanism is used to store time series input and output data collected in a predetermined forecast period; the first dynamic parameter identification mechanism is used to determine based on the non-fixed input and output data stored in the input and output data storage mechanism Parameter 値; a static parameter identification mechanism for determining constraints to be imposed on the parameters based on static input and output data; and a second dynamic parameter identification mechanism, based on non-static input and output data stored in the input and output data storage mechanism Output data, adjust the parameter 値 determined by the first dynamic parameter identification mechanism so that the parameter 値 is within the range of the constraints determined by the static parameter identification -11-(8) (8) 200404195, and the parameter is値 Spread to parameter 値 storage mechanism. According to the present invention, the use of the second dynamic parameter identification mechanism improves the prediction accuracy of a particular part of the output and enables more precise prediction. The prediction system further includes a display device for simultaneously displaying the input and output data stored in the input and output data storage mechanism and the prediction data corresponding to the input and output data and provided by the prediction model forming unit; and a marking device for The part of the prediction model required to reduce the prediction error and the cause variable that is considered to cause the error are marked; among them, the second dynamic parameter identification mechanism identifies the parameter of the cause variable marked by the labeling device to reduce the Some prediction errors. According to the present invention, the difference between the prediction 値 and the real output can be visually recognized 'and a precise prediction can be obtained, and the degree of improvement in the prediction accuracy of the part whose prediction accuracy is to be improved can be visually recognized. The prediction system further includes a display device for simultaneously displaying the input and output data stored in the input and output data storage mechanism and the corresponding input and output data and the prediction data provided by the prediction model forming unit; and adjustable parameters Marking and display device, which is used to mark the prediction model and at least two parameters for selecting the cause variables for manual tuning. For the part of the prediction model, the data displayed relative to the display device needs to be reduced. Predictive errors, and causal variables that are considered to be the cause of the errors. According to the present invention, the prediction accuracy of the part that needs to reduce the prediction error can be manually adjusted, and the degree of difference between the prediction 値 and the real output can be visually recognized. In particular, after the -12- (9) (9) 200404195 parameter is selected by the adjustable parameter selection device, only one parameter needs to be changed, so the adjustment can be easily obtained, and the prediction result due to the adjustment is in Visually recognizable. [Embodiment] First Embodiment A prediction system according to a first embodiment of the present invention will be described with reference to the drawings. Fig. 1 shows the basic configuration of the prediction system according to the present invention applied to the prediction of the inflow amount into the sewage field. It is mentioned that the inflow prediction is a specific embodiment of the operation of the prediction system and the contents described below can be applied to any prediction system. Refer to Figure 1 'The forecasting system applied to the sewage system inflow system 1 includes rainwater measuring device 1 1. Inflow measuring device 2 and rainwater inflow processing 1 3' Rainwater inflow processing i 3 represents the rainfall observation field and inflow Measure the passage of outflows and inflows between observation fields. The rainfall measuring device 11 continuously or at predetermined intervals measures the amount of rainfall at a plurality of inflow locations in the sewage field included in the sewage field inflow system 1. The inflow measuring device 12 measures the inflow into the sewage field included in the sewage field inflow system 1 continuously or at a predetermined period. As shown in FIG. 1, the prediction system of the present invention includes a prediction model device 2 and an input and output data storage mechanism 3. The prediction model device 2 includes a prediction model formation unit 21 and a parameter 値 storage unit 22, and the prediction model formation -13- (10) (10) 200404195 The unit 21 includes a prediction model, an input device for inputting data into the prediction model, and an output device for providing the data provided by the prediction model. The parameter / storage unit 22 is used to store the formation of the prediction model. The parameter 値 of the prediction model of the unit 21, the input and output data storage mechanism 3 is used to store the rainfall data provided by collecting the rainfall measurement device 11 at a predetermined cycle and the inflow measurement device 12 Time series data obtained from inflow data. The input and output data storage mechanism 3 is connected to the static parameter identification mechanism 4 and uses the data stored in the input and output data storage mechanism 3 and checks the characteristics of the sewage field inflow system 1 to determine the device to be applied to the prediction model device. The restriction condition of the parameter 値 in the parameter 値 storage unit 22 of 2 and the number of parameters deciding mechanism 5 connected to the identified parameter are used to determine the number of parameters of the prediction model forming unit 22. The dynamic parameter identification mechanism 6 is connected to the input and output data storage mechanism 3 ° The dynamic parameter identification mechanism 6 is connected to the input and output data storage mechanism 3 ° The dynamic parameter identification mechanism 6 determines the parameters of the prediction model forming unit 2 2 to be used by The rainfall data and inflow data stored in the input and output data storage mechanism 3 accurately determine the dynamic relationship between the rainfall data and the inflow data provided by the inflow system 1 of the sewage plant. The input and output data storage mechanism 3 includes an identification / confirmation data storage mechanism 31 and a prediction data storage mechanism 32. Identification / certification data storage unit 31 will identify the parameters used to construct the prediction model and confirm the accuracy of the prediction. The prediction data storage mechanism 32 stores passing data obtained in a passing period between a prediction point required for a real online prediction and a predetermined passing point. -14- (11) (11) 200404195. The operation of the first embodiment will be explained. The rainfall measurement device 11 and the inflow measurement device 12 respectively measure rainfall data and inflow data in a predetermined period of time and store them in the input and output data storage mechanism 3 as time series data. Since no prediction model is constructed at the beginning, rainfall data and inflow data are initially stored in the identification / confirmation data storage mechanism 31. Next, the static parameter identification mechanism 4 uses the time-series data stored in the identification / confirmation data storage mechanism 31 to calculate a restriction condition. A significant difference between the present invention and the conventional prediction model construction method based on system recognition lies in the use of a static parameter identification mechanism 4, which is the most important component of the present invention. The operation of the static parameter identification mechanism will be specifically explained. The effective rainfall and outflow coefficients required to determine the constraints to be imposed on the parameters of the prediction model for predicting the inflow will be explained. The effective rainfall is the rainfall that actually flows into the sewage field, which is equal to the total rainfall minus the rainfall infiltration into the surface. The ratio of effective total rainfall to total rainfall will be called the outflow coefficient. Assuming the outflow coefficient k is fixed, the relationship between rainfall R (t) and inflow Q (t) is expressed by equation (3) t—tf t «tf jQ (t) = k ^ R (t) (3) t · * 0 t »0 where t is a parameter representing time and tf is the time when the rain stops. Since rainfall is regarded as the sum of effective rainfall (inflow) and rainfall infiltration into the surface, equation (3) is regarded as an expression representing the law of conservation of water. Therefore, the ability to obtain correct and effective rainfall prediction is obviously -15- ^ (12) 200404195 is a necessary condition for an acceptable prediction model. The inflow amount prediction made by the prediction model forming unit 21 including the transfer function model of the discrete-time system will be explained. The relationship between rainfall R (t) and inflow Q (t) is expressed by expressions (4), (5), and (6). Q (t) = ^^ R (t) (4)

Aiq'1) Automatic regression part A (q-1) = l + a1q-1 + a2q-2 + ". + Anq'n · (5) Moving average part B (q ^ 1) = b1q-1 + b2q'2 + ... + binq-ni · · · (6) where q "is a shift operator representing one step behind time. Suppose R (t) is Ro (quantitative), that is, continuous For a fixed rainfall, the inflow Q (t) = Q0 (quantitative) is expressed by equation (7).。 Know the fixed rainfall

A ⑴ 1 + a! + —— + an If the constant rainfall continues, the quantity Qo is expressed by the expression (8). Q〇 — k R〇 ··· (8) Therefore,

bi + b2 + ". + bm 1 + 3.j + 3-2 + ...

Equation (9) indicates a restriction condition: The ratio of the sum of the parameters 値 of the moving average section to the sum of the parameters 自动 of the automatic regression section must match the outflow coefficient k. Since the outflow coefficient k can be checked by using the data stored in the identification / confirmation data storage unit 31, it is possible to determine the shift -16- (13) 200404195. The sum of the parameters 値 of the moving average section is relative to the automatic regression. Proportionately appropriate. The static parameter identification mechanism 4 thus determines the conditions to be imposed, and the dynamic parameter identification mechanism 6 determines the parameter 符合 that meets the limit. In general, although the restriction conditions are not easy to identify, as long as the restriction conditions are met, the outflow indicates the law of conservation of water volume. Therefore, using this limit number, the prediction irregularities predicted by the optional prediction device conflict. Therefore, although the rotation model of a discrete-time system looks similar to a black box, the constraints are constant. Although it is assumed here that the outflow coefficient is fixed, the outflow coefficient depends on the actual inflow intensity and rainfall. In this case, it is not appropriate to use a linear transfer function model that cannot be used. In this case, this can be determined by using a non-linear model with different fixed intensities of rainfall intensity. For example, the non-linear model can be Japanese Patent 1 9 9 ', Rainwater Inflow Predicting

Rainwater Inflow Predicting Method Hammerstein model. For example, if the outflow degree polynomial approximates the rainfall intensity, it can be implemented by the fixed 识别 recognition mechanism 4 in the Hammerstein model of the quadratic polynomial. The conditions used in this case and the parameters of the shifting part 値, which are in the range of the limiting conditions on the parameters, the coefficients in the dynamic parameters 値 can be regarded as any of the participation and mass conservation functions determined under the index, that is, Quality can still be expressed, but the rainfall in some predictions: Eq. (9) and know the limit coefficients set for different parameters. Application number G 1 8 66 / Apparatus and 〃 It can be considered with two inputs so that the parameter of the static parameter moving average part -17- (14) (14) 200404195 値 The ratio of the total mouth to the sum of the parameters 自动 of the automatic regression part is a simple simple constraint Different, it's a bit complicated. The above operation is an operation example of the static parameter identification mechanism 4. The dynamic parameter identification mechanism 6 can be easily implemented, and when the static parameter identification mechanism 4 operates in the following manner f, a solid prediction can still be obtained which will not reduce the prediction accuracy when there is some interference. In the f #% parameter identification mechanism 4, the formula (4) can be expressed by the formula (丨 0) in a time zone where no shift operator is used. Q (t > s-hQ (t-Ι) -a2Q (t-2) '., AnQ (t-n > + biR (t-1) + b2R < t-2) +… bmR (tm) ( 10) It is assumed that the weather is good and R (t) = 〇. Generally speaking, since most sewage plants are mixed sewage systems, and sewage containing domestic wastewater and industrial wastewater will flow into the sewage plant even without rainfall. In this case, '(ii) can be obtained from (1 0). Q (t) =-a! Q (tl) -a2Q (t-2) -...- anQ (tn) (11) Actually, 'the amount of sewage inflow changes moderately. However, for simplicity', it is assumed that the amount of sewage inflow is a fixed inflow amount qg. Then, the expression (1 1) can be written as:

Qo: (-si2 -... Q〇 (12) In order to keep equation (1 2) well, ㈤ is equivalent to the fact that A (1) is 2 0, where A (l) is the automatic regression part A (q ^ The sum of the parameters 値. This is also equivalent to the automatic regression unit A (q-1) having an integrator and the zero point of A (q) is equal to 1. -18- (15) (15) 200404195 When using a model with an integrator to make a prediction, the interference effects of fixed chirp can be invalidated. Since the sewage inflow can be regarded as a fixed disturbance, 'the impact of sewage inflow can be made completely by conforming to A (1) = 〇 Invalid 'outcome' obtains a solid prediction. However, if A (1) = 〇 'then the right side of equation (9) is infinite'. Therefore, the constraints cannot be satisfied. When A (1) is expressed by equation (1 3) This problem can be solved with very small 値. A (l) = ε, Ο < ε «1 .......... (13) Considering the stability of processing that can be checked separately, e値 must be positive. From Equation (13), it can be seen that a solid prediction can be obtained, and at the same time, the static parameter identification mechanism 4 will be reduced to a very practical condition, and the automatic regression unit The individual sums of the parameters of the moving average part are fixed as shown in the following expressions. Automatic regression part: (1) = 1 + + 3 · 2 + · · · + a. = 0 < ε ... .. (14) Moving average part: Β (1) = + t > 2 + · · · bm = s ..... (15) It is clear from equations (14) and (1 5) When the outflow coefficient k and the constant ε of the positive 値 are very close to zero, the total of the parameter 値 of the automatic regression unit and the total of the parameter 値 of the moving average unit must satisfy equations (1 4) and (1 5). The dynamic parameter recognition mechanism 6 only needs to identify the parameters in order to meet the conditions of "the sum of the parameters 自动 of the automatic regression section is fixed" and "the sum of the parameters 移动 of the moving average section is fixed", so it can achieve a very good expectation -19- (16) (16) 200404195 The static parameter identification mechanism 4 operates as described above. The number of identified parameters determination mechanism 5 determines the number of parameters η and m to be identified by the automatic regression section and the moving average section. The number η and m can be determined automatically or manually. The number of parameters η and m can be determined by the following mechanical method. The data in the existence identification / confirmation data storage mechanism 31 is divided into identification data and confirmation data. Using the identification data and the number of changes n and m, the identification parameters are identified by a generally known parameter identification method. Method. Then, the prediction is made by using the identified parameters and confirmation data '. The number of identified parameters η and m is the minimum that minimizes the sum of the squared prediction errors of the confirmed data. In general, this method is called the degree of determination method using a loss function. The optimal number of identified parameters η and m can be determined according to a criterion called (Akaike Information Criterion / Information Criterion). Instead of AIC, it can be called MDL (Minimum Description Length). As these methods determine the numbers η and m, dynamic parameters can be identified simultaneously. This identification is performed only on the determination of the optimal number of identified parameters, so that one of the identified parameters does not need to satisfy the constraints determined by the static parameter identification mechanism 4. Therefore, these 値 will not be used as the final parameter 値. Provided that one of the identified parameters satisfies the restriction conditions to be imposed on the parameters determined by the static parameter recognition mechanism 4, the recognition result can be directly used as the dynamic parameter recognition mechanism 6. The number of parameters η and m can be determined by the following manual method. Retrieve rainfall data and inflow data from the identification / confirmation data storage unit 31, -20- (17) (17) 200404195 and the rainfall data and inflow data will be plotted on the screen of the appropriate display device in time series. The rate of change of inflows in response to changes in rainfall is then estimated. If the inflow quickly responds to changes in rainfall, the effect of the automatic regression unit will increase as the number η increases, and if η is large, the change in backward rainfall is predicted, so η = 1 is preferred. Although a solid prediction can be made in the presence of fixed interference as described above, when 1 is present, the prediction is not very robust when there is variable interference. On the other hand, if the causal relationship between rainfall and inflow is only a qualitative relationship, where the change in inflow does not respond quickly to the change in rainfall, and the inflow increases with rainfall, the use of large The number as the number η is equal to the importance of placing the previous rainfall as a predictor more than the importance of placing the rainfall. Therefore, it is preferable that the number η is increased to a certain degree. When the number? Is large, when there is variable interference, a fairly robust prediction can be made. There is a qualitative relationship between these factors. Table 1 Follow the prediction of η '丨 Dependence of energy and ruggedness Rainfall following performance Anti-interference ruggedness η Fast response η = 1 is strong enough to resist fixed interference Large η Slow response is strong enough to resist fixed And the relationship not shown in Table 1 is used as a criterion for determining the number η. For example, the number η is determined in the following manner. Duration of the effect of graphs obtained by plotting -21-(18) 200404195 based on time series on rainfall data and inflow data on inflows. For example, the graph of the individual changes in inflow over time shows how much time it takes for the inflow to return to the inflow when the weather is good. Multiple rainfall events repeat the impact of rainfall on the inflow, which can determine the effect of rainfall on the inflow. The impact of the amount. For example, suppose the duration of the assessment of the impact on the inflow of 5 minutes is 30 minutes. However, if η > 1, it is smaller than the final 値 m of m 估算 which is estimated from the graph of the time series and inflow data, because when n> 0, the prediction is automatically returned. 'The above operation is the operation of the dynamic parameter identification mechanism 5. The dynamic parameter identification mechanism 6 specifically determines the number of parameters which pass the conditions of the static parameter identification mechanism determined by the decision mechanism 5.' For example, the following is to perform an exploration optimization method called genetic deduction. The following expression is defined as the evaluation function (adaptability J = XiQW-QW) = (y-y) T (y-y) ι »1

y: = [Q (i) Q (2HqM; T Among them, the text marked with A above is prediction 値, N generations wither, then, it is decided to meet the static parameter recognition mechanism 4, and the evaluation of rainfall starts from displaying the rainfall and the end of the rainfall The measurement level after the beam. The approximate duration line of the duration of the response and the rainfall shell. J m = 30/6 = 50. Displaying the rainfall data. It can be implemented as a continuous correction by the department. For example, the number of identified parameters is determined by the optimization method, and the inverse of this method can be achieved). (16) (17) (18) E number of data. Decided restrictions -22- (19) 200404195 Multiple initial parameters in the range of 値. A number of 决定 'must be decided so that the sum of and the sum of ... are fixed' This can be achieved by removing one of the parameters of each of & and h '. When ~ and bl are defined by using random numbers and in the following expressions to generate ai (i = 2, 3, ... and η) and bi (i = 2, 3, ... and m) other than ~ and bl , The total needs to be fixed

Then, the 値 generated by using a random number is represented by an appropriately sized binary notation 'and the binary 値 is used as the original individual of the genetic deduction method. By repeatedly accepting the intersection, mutation, and selection of the binary 値, it is determined 値 that minimizes the evaluation function represented by equation (16). In this way, the dynamic parameter recognition mechanism is executed. J = (yy) T (yy) e (y_4 > e) T (y —φθ) Φι Φ12. • ΘΓ Η- Φιι Φι2 · θι1) y-^ 21 Umbrella 22 θ2. Φ21 Φ · 22 Μ) Qi = ( -3 ·! / * ~ 3 · 2 '... ，-Βη] (20) ㈣ q2: = [b!' B2, ... a bn] (21) where Φ is stored in the identification / confirmation by appropriate allocation The time series rainfall data R (1) and the time series inflow data Q (t) in the data storage mechanism 31. The limiting conditions determined by the static parameter identification mechanism 4 are expressed by equations (22), (23), and (24). -23- (20) (20) 200404195 en®i = (22) d = ek (23) [U… 1,] T (24) 2 The dynamic parameter recognition mechanism 6 can be formulated with a linear equation limitation on the parameters Become a secondary planning problem. By using a Lagrange method with an indefinite coefficient, for example, the restricted evaluation function is reduced to the following new unrestricted optimization problem, and a restricted optimization method can be solved. J = (γ-ΦΘ) τ (γ-ΦΘ) + λ1 ^ θ1-1 + ε) + λ2 ^ θ2-ε ^) (25) where λ i and λ 2 are Lagrangian uncertain coefficients. By making the partial differential equation (25) 0 and make? ? ? ? ? . f = 〇 ㈣ By solving equation (26), the dynamic parameter identification mechanism 6 can be executed, and the constraint conditions determined by the static parameter identification mechanism 4 can be satisfied. For example, the solution method is described in detail in Imano Hiroshi, Yamashita Hiroshi's "Nonlinear Programming", Nikka Giren (1 994) ° After the dynamic parameter recognition mechanism 6 determines all the dynamic parameters 値, these 値 will be stored In the parameter 値 storage unit 22. After constructing a prediction model with a prediction system, the following operations are performed for prediction. The rainfall measurement device and the inflow measurement device 12 will measure the rainfall and inflow respectively in the online measurement mode at predetermined times to obtain data. The data obtained by the rainfall measurement device 11 and the inflow measurement device 12 will be stored in the prediction storage unit -24- (21) (21) 200404195 3 2 of the input and output data storage unit 3. The forecast data storage mechanism 32 will store at least m pieces of rainfall data and n pieces of inflow data. Then the 'parameter 値 storage unit 22 transmits the parameter 给 to the prediction model forming unit 2 i of the prediction model device 2, and then the rainfall data and the inflow data stored in the prediction data storage mechanism 32 2 are transmitted to the prediction model formation. Unit 2 1. The prediction model forming unit 21 forms a prediction model represented by the formula (2 7). Q (t > = --aiQ (t-1 bu a2Q (t-2) '--anQ (t-n > -HbiR (t-1) + b2R (t-2) + ... bmR (tm) (27) The rainfall data and inflow data will be given to the prediction model formed by the prediction model forming unit 21, and then the inflow is predicted by using formula (27). By repeatedly using the formula (2 7), The calculation can predict the inflow for a long time. When the final prediction is completed, the rainfall data stored in the forecast data storage mechanism 3 2 and unnecessary data in the inflow data will be transmitted to the identification / confirmation data storage mechanism 31. A series of operations are performed at a predetermined prediction period. The prediction system in the first embodiment is different from the conventional black box approximation prediction model construction method, and it uses the constraints on the parameters determined by the static parameter identification mechanism 4 to Identification of the parameters by the dynamic parameter identification mechanism 6. In this way, the prediction system of the present invention has the following effects. (1) The restriction conditions of the parameters determined by the static parameter identification mechanism 4 generally correspond to the law of physical conservation or the like, and are necessary Necessary conditions met by the predictive model In this way, it is possible to construct a prediction model that can satisfy these conservation laws with certainty. -25- (22) (22) 200404195 (2) If the data contains abnormal data or when the prediction model is constructed by the conventional prediction model formation method Insufficient data and the prediction model cannot function at all, the parameters will not diverge. The prediction model formation method of the present invention avoids this phenomenon and can form a robust prediction model. (3) No matter how the interference can obtain a solid prediction and obtain unaffected interference Prediction. Second Embodiment A second embodiment of the present invention will be described with reference to Figs. 2 to 5, wherein portions similar to the first embodiment will be denoted by the same reference numerals and descriptions thereof will be omitted. Fig. 2 shows the second embodiment. The prediction system in the embodiment is applied to the prediction of river water level. The river system 50, that is, the operation target of the prediction system shown in FIG. 2, includes a tidal water level measurement device 51, and a pump discharge measurement device 5. 2. Drain pump discharge measurement device 5 3. Drainage pump station water level measurement device 5 4 and river treatment 5 5. River system 5 0 includes regional river treatment 5 5. Causal variables that affect river water levels, and river water levels (ie, result variables). Figure 3 shows the river system 50 specifically. Referring to Figure 3, the river system 50 includes a tidal water level measurement device 51, which continuously or The tidal water level of the river water level in the drainage pumping station 5 4a that affects the river system 50 is measured at a predetermined period of time. The tidal water level measurement device 5 1 is installed in the bay connected to -26- (23) (23) 200404195 Drainage pumping station 5 4 a is installed next to the river. The upper pump discharge measurement device 5 2 included in the river system 50 will measure the river water of the drainage pumping station 54a Φ affecting the river system 50 continuously or at a predetermined time. Pump discharge of the pump above the water level. Pump emissions are affected by rainfall in this area. The river system 50 also includes a rainfall measuring device 5 8 ° The drainage pump discharge measurement device 5 of the river system 50 0 The continuous measurement or the measurement within a predetermined period directly affects the drainage pump station 5 of the river system 5 4a Φ The pumping capacity of the drainage pumping station of the river water level is 5 4 a. The prediction system shown in Figs. 2 and 3 is similar in configuration to the prediction system shown in Fig. 1, but the former is applied to the conditional prediction of the river system 'and the latter is applied to the inflow prediction of the sewage field. The prediction system in the second embodiment is different from the prediction system in the first embodiment in the following two points. One of the two points is that the prediction system includes a first dynamic parameter identification mechanism 61 and a second dynamic parameter identification mechanism. 62 instead of a single dynamic parameter identification mechanism. Another point is that the display device 7 and the marking device 8 are interactively connected to the second dynamic parameter identification mechanism 62. The operation of the prediction system in the second embodiment will be explained. The tidal water level data, the upper pump discharge data, which are measured by the water level measurement device 5 1, the upper pump discharge measurement device 5 2, the drainage pump discharge measurement device 53 and the drainage station water level measurement device 54 in a predetermined period, Drain pump discharge data and drainage station water level data will be given input and output data || storage mechanism 3 and stored in time series. Since no predictive model is constructed at the initial stage, these data will be stored in the identification / confirmation data storage machine -27- (24) 200404195 Structure 31. The first dynamic parameter identification mechanism 61 will recognize the dynamic parameters of the prediction model in a known manner. Hypothetical numbers of systems with multiple tidal water levels, upper pump discharges, and drain pump discharges were identified. However, in many cases where the number of data is insufficient, correct identification is obtained. In this case, the parameter identification is performed for each part in Japanese Patent Application / 2002 `` River Stage Prediction System ''. When the first dynamic parameter identification mechanism 61 uses the method described in the root application number 2423 1/2002 vv River Stage Predictic, it is based on, for example, the number of AIC or MDL criteria. Therefore, the number of identified parameters determines the dynamic parameters of the mechanism. The identification mechanism 61 operates simultaneously. Since the static parameters need to be determined when the first different mechanism 61 is executed, the other mechanism 4 and the first dynamic parameter recognition mechanism 61 simultaneously, when the number of data is small and the conditions for identifying the problem are poor or the materials are covered, borrow By executing the first dynamic parameter identification mechanism 6 1 the parameters may be very unreliable. In this case, the static parameter identification mechanism 4 described again. For example, since the position where the tide level is measured is below the measured position, the difference 値 L between the tide level C (t) and the water level H (t) stored in the identification / confirmation data 31. Suppose that the tidal water level system identification method is used in the prediction model forming unit 21, that is, in the following dynamic parameters, the method described in No. 2 4 2 3 1 cannot be requested according to this patent) n System 〃 to determine the parameter 5 Will operate with the first dynamic parameter and the static parameter. However, it contains abnormal data to determine the static execution. The water level is expected to be stored by the storage organization. On average, there will be a relationship between the river level and the river level. -28- (25) 200404195 is expressed by the following discrete-time transfer function model. DCq · 1 ... + dnq'n (automatic regression section) (29) N1 (q 1) = nilq 1 + ni2q 2+ ... + nimq m (moving average section) (30) When the average tide level C 〇 When continuously input, it is assumed that the average river water (31) bit is H〇TJ Nl (l) n nll + n12 +… + nlm n H〇w 〇ssi + d1 + d2ndTu So, 1 = Co · N (1) / D (1)-Co must be maintained. Therefore, the total ratio of the parameters 自动 of the automatic regression section and the moving average section is expressed by equation (3 2).

Nl (l) l-C0 Basically, for 以外 other than C (t), the formula (3 2) must be good. However, it is practically impossible. Therefore, the ratio between the sum of the numbers determined by the identification of the first dynamic parameter recognition mechanism 61 and the difference between the ratio and the ratio shown in equation (3 2) is different. Then, the formula (32) is stored as the static parameter identification mechanism 4, and the state parameter identification mechanism 62 will identify the parameters again to meet the restriction conditions determined by the static parameter identification mechanism. Similarly, the relationship between the discharge amount P (t) of the drainage pump and the river water level H (t -29- good (32) maintaining the watch's reference when the second action number) (26) 200404195, and the upstream pump The relationship between the discharge amount Pu (t) and the river water level is maintained by the static parameter identification mechanism 4 when needed. When the discharge pump discharge amount P (t) and the upper pump discharge amount pu (t) are maintained in advance, the river water level will not be stable. If the upper pump discharge Pu (t is fixed and the drainage pump discharge P (t) = 0), the water level will continue to rise unless the river floods. If Pu (t) = 0 and p (t) are solid, the river water level will continue It is reduced until the river dries up. Therefore, the relationship between the recognized discharge amount and the river water level has an integral feature. In this case, the following static parameter identification mechanism 4 is possible. For example, if the pump discharge amount Pu (t) is If it is fixed, the river will rise at a fixed rate. The rise of the river water level relative to the unit discharge ratio (gradient) s (t) can be expressed by the following expression. H (t) determines when the ° is fixed, otherwise the river is fixed, For the pump. S (t) Pu (t) N2 (q- oiq · 1) (lq-1) N2 (q) Pu (t) -you W) pu (〇 (33) (34) N2 (q -1) = where 'N2 (q · 1) is a polynomial composed of a system for pump discharge Pu (t). If this river system is similar, the relationship has an integral characteristic life) = (w) Bingyi 】), Must be kept good. Substitute this relationship into the formula (, and obtain the formula (3 5). For the number parameter, then 33) -30- (27) 200404195

Pu (0 s (t) Pu (t) «N2 (q-1) If pu (t) is two pu 〇 (constant), the rising rate of this fixed bit is consistent and converges with fixed 値. N2 (l) is below ^ D (q-) = -1 D (q-) + (1-q-) 0 (1) = q ~ l = -1 ((^ + 2 (^ +… + ndn) N2 (l) n21 + n22 ten · •• eleven_ S = ϋ (1) = two 1 ♦ + 2d2 + · · · Equation (3 9) indicates the restriction conditions to be added to the parameters of the static parameter recognition machine. Because s is a gradient representing a curve having a change in unit water level. Therefore, the true gradient of time in one step when the discharge of the pump is Pu0 is set to S. The coefficient identification mechanism with integrated characteristics can be identified in the above-mentioned manner. 4 .If the first dynamic parameter identification (35) is used, the river water calculated with the river water (36) (37) (38) (39) structure 4 must be calculated by dividing the relative by Pu0 to perform static When the parameter on the right side of the formula (3 9) calculated by the parameter determined by the participating institution 61 is -31-(28) (28) 200404195, and the difference between s and s is large, the second dynamic parameter identification mechanism 62 will identify the parameter again. Satisfying (39) ° will explain the second dynamic parameter identification The identification operation of the structure 62. As shown in FIG. 2, the second dynamic parameter identification mechanism 62 is connected to the display device 7 and the marking device 8. The display device 7 is also connected to the first dynamic parameter identification mechanism 61. Figures 4 and 5 The display device 7 and the labeling device 8 are illustrated as an example. Referring to FIG. 4, when the first dynamic parameter identification mechanism 61 is used to identify the dynamic parameter, the predicted river water level can be displayed by the display device 7. Meanwhile, stored in Identification / confirmation of upper pump discharge data, drainage pump discharge data, tidal water level data and river water level data in the data storage device 31 can be displayed by the display device 7. It can be easily understood from FIG. 4 that the display device 7 The displayed data has a high prediction accuracy part and a low prediction accuracy part. The labeling device 8 separates the low prediction accuracy and the prediction accuracy that needs to be improved as shown in FIG. 4. At the same time, the labeling device 8 Infers the variable that is considered to be the largest cause of low prediction accuracy (river water level, tidal water level, upper pump discharge or drainage pump discharge) and displays the cause variable. Example In other words, in Figure 4, when the discharge from the upper pump and the discharge from the drain pump are substantially zero, the prediction accuracy is satisfactory; that is, when there is no rainfall, the prediction accuracy is satisfactory. Therefore, it can be known that rainfall is the reason for reducing the accuracy of the forecast. Because the tidal water level does not depend on the rainfall at all, the tidal water level will not be the reason for the decrease in the accuracy of the prediction. -32- (29) (29) 200404195 The prediction accuracy is quite high in the state where the pump discharge is increasing. 'The pump discharge is hardly the cause of the decrease in prediction accuracy. In a state where the discharge capacity of the drainage pump suddenly starts to increase, the prediction accuracy is not acceptable. As a result, it was inferred that the reason for the decrease in prediction accuracy was the coefficient of the river water level (automatic regression coefficient) and the secondary reason was the drainage pump discharge. In this case, the 'river water level is considered to be the main cause and the marking device 8 indicates the river water level as the cause. Therefore, the labeling device 8 labels the portion where the prediction error is desired and the cause of the prediction error. The parameter of the river level specified as the cause is then stored and backed up. Then, the second dynamic parameter recognition mechanism 62 performs recognition to reduce the prediction error. Specifically, the second dynamic parameter identification mechanism 62 optimizes the parameter 値 using a search optimization method such as a genetic deduction method or a restricted quadratic programming method. The second dynamic parameter identification mechanism 62 differs from the first dynamic parameter identification mechanism 61 in that the evaluation function is defined not to be used for all prediction errors but only for prediction errors in the portion marked by the device 8. And the dynamic parameters are identified so as to reduce the prediction error in the marked part. Therefore, in some cases, when the second dynamic parameter recognition mechanism 8 recognizes the parameters again, the prediction error in a portion other than the portion marked by the marking device 8 may increase. The homotopy approximation method can be used to avoid this phenomenon, which uses the parameter 辨识 identified by the first dynamic parameter recognition mechanism 61 as the initial 値 and gradually and continuously changes the initial 値. If the gradually changing predicted parameters are simultaneously displayed by the display device 7, then when the approximate predicted parameters are displayed by the display device 7, the operator performing the parameter identification can stop the second dynamic parameter identification-33- (30) ( 30) 200404195 Parameter identification operation of mechanism 62. If the method is executed by the second dynamic parameter identification mechanism 62 and the parameters cannot be identified satisfactorily, the parameter of the river water level identified by the first dynamic parameter identification mechanism 6 1 for its original value and another reason variable will be determined by The marking device 8 marks. For example, drainage pump discharges are labeled as cause variables. Then, the second dynamic parameter identification mechanism 62 executes the same method to identify the parameters. In this way, the prediction error in the portion marked by the marking device 8 will be reduced as shown in Fig. 5, and the prediction accuracy of other portions can be improved. In this way, the prediction model in the prediction mode forming unit 21 is configured. Next, in accordance with the method described in the first embodiment, a prediction is obtained. As is clear from the above description, the prediction system provided with the first dynamic parameter identification mechanism 61 and the second parameter identification mechanism 62 can improve the prediction accuracy of a particular part and can provide a prediction model with strict prediction. Third Embodiment A prediction system according to a third embodiment of the present invention will be described with reference to Figs. 6 and 7. The prediction system in the third embodiment is substantially the same in configuration as the prediction system in the second embodiment shown in FIGS. 2 to 5, but differs from the prediction system in the second embodiment only in that the The parameter selection device 9 is adjusted instead of the marking device 8. In the prediction system shown in Figs. 6 and 7, the second dynamic parameter recognition mechanism 62 performs manual dynamic parameter recognition instead of automatic dynamic parameter recognition. -34- (31) (31) 200404195 The coefficient parameter of the cause variable as shown in Fig. 7 is displayed in the part of the display device 7. Select the cause variable that is estimated to reduce the accuracy of the prediction in a particular place. Select as many pairs of parameters as the cause variable to be adjusted. For example, the coefficient parameters n31 and n34 of the drainage pump discharge volume shown in FIG. 7 are selected for adjustment. Select the two parameters according to the following criteria. In the table shown in Fig. 7, with respect to time prediction, the farther right parameter represents the farther past influence, and the farther left parameter represents the closer to the past. In the table shown in Figure 7, the shorter the distance between the two selected parameters, the greater the impact at a particular time, and the longer the distance, the longer the duration of the impact. The adjustable parameter selection device 9 refers to the data displayed on the display device 7 and selects two parameters. Since the adjustable parameters n31 and n34 are selected, these parameters will be adjusted in pairs. The static parameter identification mechanism 4 sets a restriction condition that the total of the parameter 値 is required to be consistent with the constant. Therefore, when determining 之一 for one of the two parameters, 値 for the other parameter is automatically determined. For example, if 値 of parameter n31 changes from -1.2 to -0.8, 値 of parameter n34 must change from 0.6 to 0.2. In this way, the parameter 値 can be adjusted to meet the limiting conditions determined by the static parameter identification mechanism 4. When the parameter adjustment operation is performed, the data displayed on the display device 7 is observed visually by predicting the change in the prediction caused by the parameter adjustment. Taking the standards (a) and (b) into consideration, gradually adjusting the parameters, and the display device 7 will display -35- (32) (32) 200404195 as shown in Figure 5 for the final prediction 値. As can be seen from the above description, in the prediction system of the third embodiment, 'by connecting the adjustable parameter selection device 9 to the second dynamic parameter identification mechanism 6' to adjust the two parameters in pairs, the actual parameter One of the fine-tuning can artificially improve prediction accuracy. Therefore, when it is necessary to improve the accuracy of condition prediction in a specific place, the artificial system can obtain parameter fine-tuning. Although it is believed that it is difficult to adjust the parameters by the black box approximation because the parameters have no physical meaning, the present invention can still fine-tune the parameters even when the black box approximation is used. The present invention has the following effects. The determination of the static constraint conditions, that is, the constraint conditions of the parameters in the steady state, will be able to identify the dynamic model by a combination of static model identification and dynamic model identification under the constraint conditions. Therefore, prediction models using black box approximation can take physical constraints such as the law of conservation of mass into account. Even under the unfavorable conditions that the available data includes abnormal data or insufficient data, a reliable prediction model can still be established. Using artificial discrimination, fine-tuning of the parameters of the prediction model using the black box approximation can be obtained to improve prediction accuracy. [Brief Description of the Drawings] Figure 1 is a block diagram of a prediction system according to the first embodiment of the present invention. Figure 2 is a block diagram of a prediction system according to the second embodiment of the present invention. -36- (33) (33) 200404195 Figure 3 is a view of a river system; Figure 4 is a display device and a marking device before adjustment; Figure 5 is a display device and a marking device after adjustment; Figure 6 is a block diagram of a prediction system according to a third embodiment of the present invention Figures 7 and 7 help explain the operation of the adjustable parameter selection device. The main elements: pairs of items! Treatment table 1 Sewage field inflow system 2 Prediction model device 3 Input and output data storage mechanism 4 Static parameters 3 ijl 5 Different institutions 5 After 5 different parameters, the decision mechanism 6 Motion m Parameter 5 Moebei mechanism 7 Display device 8 Marking device 9 Mita parameter selection device 11 Rainfall measurement device 12 Inflow measurement device 13 Rainfall-inflow processing 3 2 1 Forecast model forming unit 22 Forecast model forming unit 3 1 m donation / confirmation data storage mechanism 32 forecast data storage mechanism-37- (34) 200404195 50 river system 5 1 tide water 52 upper pumping station 52a upper pumping station 53 drainage pump 54 drainage pump 54a drainage pump 55 riverside 5 8 rainfall量 6 1 第 — ^ 动 62 Second position measurement device Displacement measurement device Station emission measurement device Emission measurement device Station physical measurement device State parameter identification Mechanism state parameter identification Structure • 38-

Claims (1)

(1) (1) 200404195 Patent application scope 1 · A prediction system for predicting a prediction target, including: a prediction model forming unit for predicting the prediction target, including a prediction model with multiple parameters; a parameter 値 storage A unit for storing multiple parameters; an input and output data storage mechanism for storing time series input and output data for prediction; a static parameter identification mechanism, which determines the parameters to be added to the parameters based on the fixed input and output data Restriction conditions; and dynamic parameter identification mechanism, based on the non-fixed input and output data stored in the input and output data storage mechanism to determine the parameters 値 that meet the limits determined by the static parameter identification mechanism, and transmit the parameters 値 to Parameter 値 storage mechanism. 2 · The prediction system according to item 1 of the patent application scope, wherein the prediction model forming unit includes an automatic regression section and a moving average section, and the automatic regression section linearly uses the time series output data as a prediction model, and the moving average section performs time series The input data acts linearly; and the static parameter identification mechanism imposes a limiting condition on the parameter 値, so that the ratio of bb / aa is fixed, where aa is the sum of the coefficients 自动 of the automatic regression section, and bb is the moving average section. The sum of the coefficients 値. 3. If the prediction system of the second item of the patent application scope, the static parameter identification mechanism will impose restrictions such that the total aa of the parameter 値 of the automatic regression unit is a fixed positive number very close to zero and aa = e represents ε < < 1. -39- (2) (2) 200404195 4 · For example, the prediction system of the second scope of the patent application, in which the dynamic parameter identification mechanism, the static parameter identification mechanism, according to the requirements, use the optional optimization mechanism and stored in the input and The non-fixed input and output data in the output data storage mechanism must be fixed in the previously determined parameters of the automatic regression section and the moving average section. The parameter 値 must be fixed. 5. · A kind of prediction system for predicting the prediction target, including: a prediction model forming unit for predicting the prediction target, including a prediction model having multiple parameters; a parameter 値 storage unit for storing multiple parameters 値; input And output data storage mechanism for storing time series input and output data collected for prediction during a predetermined period; a first static parameter identification mechanism, based on non-fixed input and output data stored in the input and output data storage mechanism, Determine the parameters 値; the static parameter identification mechanism, based on the real or fictional fixed input and output data, determines the constraints imposed on the parameters; and the second dynamic parameter identification mechanism, based on the non-fixed stored in the input and output data storage mechanism Input and output data, adjust the parameter 値 determined by the first dynamic parameter identification mechanism so that the parameter 値 is within a range that meets the limit conditions determined by the static parameter identification mechanism, and transmit the parameter 给 to the parameter 値 storage mechanism. 6 · If the prediction system of the scope of application for patent No. 5 further includes: a display device for displaying the input and output data stored in the input and output data storage mechanism at the same time and corresponding to the input and output data and from -40- (3) (3) 200404195 The prediction data provided by the prediction model forming unit; and a marking device for marking the part of the prediction model that needs to reduce the prediction error, and the variable considered to be the cause of the error; of which, the second dynamic The parameter identification mechanism will adjust the parameter of the cause variable marked by the marking device to reduce the prediction error in the portion marked by the marking device. 7 · If the prediction system of the scope of patent application No. 5 further includes: a display device for simultaneously displaying the input and output data stored in the input and output data storage mechanism and a unit corresponding to the input and output data and formed by a prediction model Provided prediction data; and adjustable parameter marking and display devices, which are used to indicate the part of the prediction model whose prediction error needs to be reduced relative to the data displayed by the display device, and the variables that are considered to cause the errors, and For selecting at least two parameters of the cause variable for manual adjustment. -41-