WO2008003105A1

WO2008003105A1 - Method and system for the automated determination of optimized forecasts

Info

Publication number: WO2008003105A1
Application number: PCT/AT2007/000317
Authority: WO
Inventors: Hubertus Hofkirchner
Original assignee: Hubertus Hofkirchner
Priority date: 2006-07-03
Filing date: 2007-06-27
Publication date: 2008-01-10
Also published as: DE112007001505A5

Abstract

The invention relates to a method and a system (1) for the automated derivation of optimized forecasts based on individual forecasts, for the control or regulation of operative systems and/or processes (13), wherein individual forecasts supplied by forecasting units (2.i), said units describing a value for a defined future system or process status and the statistical distribution thereof and having different weightings and noise factors, are analyzed for contradictory statements, wherein contradictory individual forecasts are eliminated and the remaining individual forecasts are accumulated into an optimized overall forecast while taking the noise factors into consideration.

Description

Method and system for automated determination of optimized forecasts

The invention relates to a method for the automated determination of optimized forecasts on the basis of individual forecasts for the control or regulation of operative systems or processes.

Furthermore, the invention relates to a system for the automated derivation of optimized forecasts for the control or regulation of operative systems or processes.

Numerous control techniques in operational systems, e.g. In industrial manufacturing processes, in power generation plants or in the control of hydrological plants, heating and irrigation systems to financial trading systems are based on automatic units for generating forecast data for certain system state or process parameters. The accuracy and reliability of such forecast data is often an essential prerequisite for a functioning control.

From DE 197 32 295 A1 it is known, for example, to use prognosis data of a single forecasting unit for a prognostic control of a heating system, wherein in particular weather data is fed to a central weather station. However, only the transmission of forecast data from a central forecasting unit is provided, whereby in the case of inaccurate forecasts an inadequate regulation of the heating system is the consequence. A similar system is described in GB 2 309 567 A, in which case additionally the return transmission of measurement data of the heating system to a central location is proposed.

The implementation of forecasting models for individual forecasts is carried out on the basis of known statistical methods and models, in particular using modern methods, such as the SOM method described in WO 2004/029738 A1 (SOM - Seif Organizing Maps - Self-Organizing Cards) or methods from the field of artificial intelligence (eg neural networks, genetic algorithms).

However, many of these methods are rarely used in automated systems because their efficiency and stability in general can not be guaranteed. One reason for this is the lack of statistically validated statements about the limits of the efficiency and validity of black box models, i. in problems related to the predictability and reliability of forecasts.

It is therefore an object of the invention to provide a method and a system for the derivation of optimized forecasts, which is based on several individual forecasts, which are evaluated to achieve a high reliability to a total forecast. The invention is based on the finding that contradictions can be eliminated when evaluating a plurality of individual forecasts over weightings and that an optimized overall prognosis can be derived from the individual forecasts, which also becomes more and more exact successively by corresponding correction of the weights assigned to the forecasting units. At the same time, a noise factor is determined which describes the validity of the optimized overall forecast.

To achieve this object, the invention provides a method and a system as specified in the independent claims. Advantageous embodiments and further developments are the subject of the dependent claims.

The present technique is based on the application of an adaptive method, with a compression coding, wherein in a first step, a plurality of forecast units transmit their forecasts for a specific system or process state to a central aggregation unit. In addition to the respective forecast value and its statistical distribution, a weighting and furthermore a noise factor are included in the prognosis signal dataset. This results in a uniform abstraction of the data of different forecast units, which is given by forecast horizon, forecast value, its statistical distribution, its forecast weight and noise factor. This abstraction is essential for the uniform transmission and return of Transmission of forecast signals of any kind suitable.

To illustrate the forecast signal record, the following example may be cited: A heavy rain forecast unit predicts a rainfall intensity in a particular catchment area for a given time with e.g. 0.35 mm / h, log normal distributed with a standard deviation of e.g. 5.2%. This prognosis is made with a weight of e.g. 172% and a noise factor describing the estimated reliability of the prognosis of e.g. 2.4% to a central processing unit.

In itself, DE 195 37 850 A discloses a method for encoding weather-related prognosis data, which, however, allows a purely meteorological and not universally applicable prognosis signal coding and also provides no weighting and no noise factor for aggregation and linking of signals of several prognosis units.

The mathematical operations described below in more detail for the elimination and aggregation of such coded prognosis signals and for the substitution and evaluation of prognosis conflicts represent preferred possibilities for achieving the desired, optimized prognostic determination, in which the weighting and the noise factor are used to aggregate and link signals of several prognosis units are or allow this.

The forecasts of two forecasting units eliminated due to contradictions are stored in a database temporarily, i. up to the end of the forecast horizon, in the form of one forecast conflict record each, which represents as a uniform abstraction the calculated forecast separation value, the calculated weighting ratio (ratio) and the overall weighting. This makes it possible, despite a plurality of synchronously or asynchronously arriving different individual forecasts at any time to determine only a single valid, iteratively improved overall forecast.

As a result, below is a capacity-saving storage and any time transmission of a single, the aggregated and current state of any number of individual forecasts descriptive record - in the manner of data compression - possible. This is particularly advantageous if a forecast also serves as an input factor for the modeling of other forecasts dependent on this value (eg temperature and precipitation) or for an iterative improvement of the forecast itself.

The present technique also allows the compression of forecasting information for storage and transmission purposes, completely independent of the respective forecasting method or mode of the associated forecasting units. This is a further difference to known prognosis determinations, which only pursue the goal of ascertaining the desired prognoses from an often large total data quantity in a specific, arbitrarily complex prognosis procedure.

The application of weights (weightings) is advantageous because the available weight of a better forecasting unit can be continuously increased, as a result of which its forecasts can be taken into greater account in the further aggregation, thus constantly improving the aggregated overall forecasts.

It should be noted that US Pat. No. 6,036,349 already discloses the use of weights, but for the validation of prognosis models, where an assessment of the quality of the prognosis units itself is aimed for, without the weight being taken into account in obtaining better prognosis information or the weight allows an automatic regulation of the influence of each prediction unit on an overall prognosis.

The present application of computational weights also provides a beneficial crosslinking effect on a plurality of forecasting units and forecasting values. Namely, if a forecast value represents an input for the forecasting model of a second forecasting unit, the latter can provide an assumed forecast value, as well as a weight and a lower noise factor, with the result that other, subordinate forecasting units can be improved by an improvement of that first forecast value. gain weight and an iterative optimization starts automatically. As a result, there is an ongoing reduction in the noise factor of all affected forecast values until finally the overall system reaches equilibrium. The use of weights thus enables an overall system of networked forecasts which is improved compared to individual forecasts.

As a consequence of the system according to the invention, it is therefore possible for several forecast units to network with respect to several forecast values, by using arbitrary methods (such as neural networks or regressions) for the determination of individual forecast values which optimize the forecast quality at the node level and optimization methods at the network level for the transmission of forecasts for the aggregation, which altogether maximize the weight allocation to the prognosis unit.

To illustrate the advantages of the present network of individual forecasts, the following example is given: A reservoir serves both for energy generation and for flood protection and irrigation. Several forecast units produce relevant prognosis information, eg for temperatures, precipitation amounts and soil moisture of the associated water catchment. The purpose of a lock control unit is to maintain an optimal (maximum) water level for power generation and irrigation on the one hand, and for the flood defector, on the other hand, taking into account the predicted inflow line, which depends on the above meteorological forecasts, on the physical characteristics of the respective catchment area is dependent on the controls of upstream hydrological facilities, but the control in turn provides prognosis data concerning the inflow course of downstream hydrological facilities (reservoirs). At the same time, meteorological (and other) forecasting data also affect forecasts for energy and water consumption, with, for example, energy consumption rising at lower temperatures to some extent, and water consumption at higher temperatures increasing in seasonal dependency. It thus becomes clear how through the networking of the different outcomes will allow for improved overall governance in terms of the three control objectives (flood protection, energy production, irrigation).

In the present technique, the aggregation of the individual predictions takes into account the noise factor transmitted by the forecast units, which must be overcome. Below is shown in more detail how the density function of a lognormal distributed forecast value is reduced by the noise factor. The noise factor limits the occurrence of a prognosis contradiction; without him any new prognosis would immediately contradict a previous one; with a noise factor, a plurality of limited different predictions may remain active. For example, a first forecasting unit for a given future point in time will require a water level of the aforementioned 8.04m hydrological facility with a standard deviation of 1.4%. This prognosis is coded and transmitted with a weight of 0.12 and a noise factor of 2.7%. On the other hand, another second forecasting unit expects a water level of 8.11 meters with a standard deviation of 1.8% and reports this forecast with a weight of 0.24 but a noise factor of 7.6%. In the same way forecasts for temperatures, precipitation, soil moisture, etc. can be coded.

A prediction discrepancy taking into account the noise factor is only present when the inverse probability increased by the noise factor (hereinafter referred to as "ratio") that the final value at the prediction time is smaller (or greater) than a certain data point, is greater than the inverse ratio (also referred to as anti-ratio) of another prediction at the same data point This can be calculated by, as will be explained in more detail below, using methods of the curve discussion, determining a zero point of the second derivative of the formula for the ratio quotient (below), According to the invention, such contradictory predictions are eliminated depending on their weights and, as mentioned above, a forecast conflict record is generated as an example of a forecast conflict for the above. Staube- The assumption is that the forecast separation value is at a level of 8.0786 with a ratio of 1.42095 (see calculations below). The weight of the first forecast of 0.12 is completely eliminated with a ratio of 1.62036, the second forecast remains active and subsequently flows into the aggregated forecast with a weighting of 0.1656, as of originally 0.24 weight are below 0.0744 (0.12 times 0.62036) in the forecast conflict bound to the forecast horizon.

For the forecast horizon, the evaluation unit assigns the total weighting of each forecast conflict data set to the forecast unit based on the actual value measured by one measurement unit or on the basis of a temporally downstream total forecast value of another aggregation unit for the same forecast value with respect to the calculated forecast separation value. The second method is used when the actual value at the forecast time in question can not be measured. In the above example, the level can be measured, for example, it could be 8.11 m, which means that the individual forecast of the second forecast unit will have more weight in the future.

The forecast units or generally input units thus regularly use a (arbitrary own) method to determine forecasts and transmit them to the system, i. to the computer resources. The system prefers returned aggregated forecasts (both at the same and other inputs) and their noise factor can serve the forecasting units in the sense of an adaptive method for the iterative improvement of their own forecasts.

The incoming forecast signals are stored by the system in a memory unit and then go through a prognosis signal processing, which first determines whether the new prognosis signal contradicts an already stored active prognosis and to which active prognosis the greatest contradiction exists. Conflicting predictions are stored by an elimination order according to the extent of the contradiction in a storage unit for forecast conflicts with the weight eliminated, and the elimination is noted in the prognosis signal data sets weight-reducing. Predictions with zero weighting become inactive. From the active forecasts remaining in the forecast memory, an optimized overall forecast is calculated in the aggregation unit by means of statistical aggregation.

The optimized overall prognosis is then transmitted to a control unit, which in turn forwards the resulting control signals to a process unit. In addition, the calculated overall forecast can also be actively queried in suitable periodicity by those forecast units which either forecast this value or need it as input, without having to transmit each individual aggregated value.

Records stored in the prediction conflict memory are evaluated at the forecast horizon by comparing the value actually measured by a measurement unit or a downstream overall forecast with them. The resulting released weights are assigned to the corresponding forecast unit.

Those forecasting units to which forecast conflict records are attributed may at any time request to substitute the weight of these conflict records against weighting of other active forecasts. The released weight is immediately available to the relevant forecast unit for new forecast signals.

It should be noted that in the prior art various improvements of individual, specifically selected prognosis models have been proposed, such as in DE 101 27 790 Al, which proposes an adaptive model that a single, predetermined prognosis model using measurements by means of a adapted correction factor adapted to counteract various types of errors modeling without having to form a new model. An advantage of the present invention in comparison is that several different models can be performed in parallel and so an automatic adaptation of the control is made possible even with less stable or variable processes. DE 197 53 034 A is based on an improvement of forecasts, in particular for traffic control processes, in that in a probabilistic selection a basic prognosis for a parameter is investigated and improved as a function of a measured value for a second parameter by the first parameter being time-dependent the second is coupled. Furthermore, old couplings are eliminated by the passage of time. By contrast, the present invention aggregates the independent forecasts of several different forecast units for one and the same basic parameter without the need to take into account any possible upstream and downstream forecasts in the network. (The procedures and input data used by the individual forecasting units are entirely open to them.) The elimination of forecasts also takes place, not simply by the passage of time, but only by the occurrence of contradictory forecasts, whereby the system according to the invention is always in an updated state of networking.

DE 601 11 238 T2 describes a process control for time-variant, non-linear processes, which triggers a corresponding switchover to another line model predictive control model in the case of a change detected by measurement between a set of several previously known states (with elimination of implausible measurements). In contrast, the generalized control according to the invention offers the advantage that a prior knowledge of possible models and a predefined model transition is not required; It is also possible to add further forecasting units with new strategies, for example during ongoing operation.

Furthermore, the literature describes adaptive forecasting methods for tax purposes in which continuous forecasts are produced for unobservable measured values, see, for example, the article entitled "Time Series: a complementary technology to control charts for monitoring analytical systems" by M. Pilar Callao and A. Rius, however, also describe only adaptations within a single prognosis model by previously eliminating erroneous observations by determining autocorrelations Therefore, it relies only on improving the performance of a particular forecasting model, but it does not offer the opportunity to constantly adapt an aggregated forecast from the results of a plurality of forecasting models of different nature.

The invention offers the advantage of networking a plurality of independent, generalized predictions, independent of the underlying model that generates them. This also has the advantage that process control can firstly take place without a restrictive preselection with respect to a specific prognostic model, and secondly, with subsequent availability of better prognosis models, the process control fundamentally improves instead of providing only a presumed error correction type for a predetermined model.

The invention will be explained below with reference to particularly preferred embodiments, to which it should not be limited, and with reference to the drawings. In detail in the drawing:

1 is a block diagram of a system for the computer-aided derivation of optimized overall forecasts according to the invention;

FIG. 2 shows an adjusted density function for a predicted water level taken as an example; FIG.

3 is a flowchart showing the procedure in the method for deriving an overall forecast from individual forecasts, with a system according to FIG. 1;

4 shows in a diagram the two ratio quotients for the example of the predicted water level, to illustrate the conflict zone determination; and

Fig. 5 in a diagram Ratiosumme about water level two ratio sums, to illustrate the Trennwertbestimmung.

FIG. 1 shows a system for the automated determination of op- timed forecasts, this system 1 receives from individual forecasting units 2.1, 2.2 ... 2.n (generally 2) individual forecasts. In the example shown, the system 1 has computer means 3 specified with a border, which have a memory unit 4 for the prognosis signals, ie individual forecasts, as well as an associated prognosis signal processing unit 5.

More in detail, the forecasting units 2.n (n = 1, 2...) Transmit forecast signals in the form of a forecast dataset containing the designation of the predicted value, the forecast horizon, the forecast value and its statistical distribution, as well as the weighting and the Noise factor represents. This results in a uniform abstraction of the data of different forecasting units 2.n, which are stored in the memory unit 4 for the subsequent signal processing in the processing unit 5.

Is turned off in the prediction units 2.n on a forecast value with the target variable X, which is a random variable in a statistical sense, and their possible variants X ₁ are, for example, a lognormal distribution (with mean μ and standard deviation σ). The adjusted density function f (x _± ) is given by:

where r denotes the noise factor, σ the standard deviation of the prognosis value, X ₁ the expression value, and μ the mean value. This function f (X ₁ ) can also be written in simplified form as:

1 + r

where LN 'denotes the log normal density function (as the first derivative of the log normal distribution function) which is adjusted by the noise factor r. It is shown in FIG. 2 with the curve 6, together with the standard density function, curve 7, shown. Specifically, FIG. 2 focuses on the example of the level forecast of the second forecast unit 2.2 explained in the introduction, with the adjusted density function 6 distributed lognormally at X = 8.04 m, with a standard deviation σ = 1.4%, and a noise factor r = 2.7%, illustrates

From this, two ratios R _m , R _m ^c can be determined in the processing unit 5 per prediction as the inverse probability of a (measurement) value above and below a certain expression value X ₁ increased by the noise factor r:

l + φ

-φxLN _m (x ₍ )

where R _m denotes the ratio and LN the lognormal distribution and the two cases of the characteristic direction are determined as: 1 measured value over characteristic value X ₁ -1 measured value over characteristic value X ₁

The complementary ratio or anti-ratio for further processing or for purposes of comparison with other forecasts is accordingly calculated as:

Each new forecast signal is checked in the predictive signal processing unit 5 to see if it contradicts an active forecast (ie, a forecast that has already been received, stored, but not yet eliminated), in which case the forecasts will be in order of greatest contradiction with the aid of an elimination unit 8 (see Fig. 1) connected to the processing unit 5.

This process may, for example, be done in detail as shown in FIG. 3: After a start step 8.1 in FIG. 3, it is first checked according to field 8.2 whether active forecasts are already stored in the memory unit 4 for the designated prognosis object of the prognosis signal. If not, no elimination is required and can be continued immediately with the aggregation in an aggregation unit 9 (see Fig. 1) also connected to the processing unit.

If already active forecasts exist, then, according to block 8.3, by a suitable method of curve discussion, such as the Newton's method or the secant method, the maxima of the two ratio quotients hMxi) (ie one quotient quotient - h ₊ , h_ - per expression of Φ [-1, + I]) between the new forecast signal and each individual active forecast.

Thus, two ratio ratios are determined for each pairing by first checking each of the two ratios of the new forecast based on the respective opposing anti-ratio (put-call or call-put) of each active forecast to a maximum. The probability distribution may vary depending on the type of the predicted value; however, for the sake of simplicity, a specific lognormal distribution typical of hydrological phenomena (e.g.

For example, the determination may be made using the following condition, wherein for the first derivative of the following condition for the ratio quotient, a root must be searched for a maximum (by a suitable method, such as the Newton's method or the secant method, as mentioned above):

n \ ^X i) - _{c /} s

where φ = [-1, + I], thus

where h is the ratio quotient at the value Xj _. is expressed as h_ at φ = -1 and expressed as h ₊ at φ = +1.

If such a maximum, with one of the two ratio quotients over unity, cf. also the query field 8.4 in Fig. 3, so there is a contradictory forecast signal, and it must by means of a prognosis conflict data set first prognosis weight be eliminated, s. Block 8.5 in FIG. 3.

4 shows an example of the two ratio quotients h and h ₊ with a predominant level forecast of Xi = 8.04 m, O ₁ = 1.4%, ri = 2.7% and a new level forecast of X ₂ = 8.11m, O ₂ = 1 _/ 8%, r ₂ = 7.6%.

In the example it becomes clear that between the new prognosis of a prediction unit, eg the prediction unit 2.2, of 8.11 m (σ = 1.8%) and the prevailing prognosis of a prognosis unit, eg the prediction unit 2.1, of 8.04 m (σ = 1.4%) in the value range of 7.94 to 8.39 m (points A, B), where the curve for the ratio quotient for the expression Φ = +1 is above the threshold of one, there is a prognosis contradiction, which is the Noise factor of 2.7 and 7.6%, respectively, since, according to the new prognosis, an expression value X ₂ up to 8.04 has a probability of 31.5% (noise-adjusted 36.34%), according to the prevailing prognosis an expression value X ₁ over 8, 04 but has a default probability of 50.0% (noise adjusted 48.69%), resulting in a ratio of 1.3396 (48.69 / 36.34). Thus, for example, 100% of the weighting of the forecasting unit 2.2 against 60% weighting of the forecasting unit 2.1 could be offset, and therefore a higher future forecast weighting exceeding the noise factor would be expected for both forecasting units 2.1, 2.2 a priori (in which the sought-for contradiction lies):

1st forecast: [60% + 2.7% = 61.6%] <[160% x 50.0% = 80.0%]

Forecast: [100% + 7.6% = 107.6%] <[160% x (1 - 31.5%) = 160% x 68.5% = 109.6%]

In this example, the second prediction unit 2, based on its own prognosis values, has a weight gain of 9.6%, which thus exceeds the 7.6% increase due to the noise factor.

In this range of values, an arbitrary characteristic value Xi can now in principle be assumed as the prognosis separation value, and for this value the respective ratios R _m , R _m ^c can be calculated, which determine the weight contribution ratio in the forecast conflict data set.

In this case, it has proven to be advantageous to determine that characteristic value X ₁ as a prognosis separation value at which the sum of the forecast weight eliminated by the prognosis conflict is the maximum. This expression value is at that value, at which in each case the respectively associated one of the two ratio sums is minimal, whereby the main condition remains that the ratio quotient must be greater than 1 at the determined separation value. This minimum ratio is calculated by looking for a minimum for the first derivation of the following formula for the ratio by one of the above-mentioned methods of the curve discussion:

where g denotes the ratio sum at the value X ₁ , expressed as g_ at φ = -1 and expressed as g ₊ at φ - +1.

An example of the two ratio sums g ₊ , g_ as above at a level Xi = 8.04 m, O ₁ = 1.4%, ri = 2.7% (prevailing forecast), or a level X ₂ = 8, 11 m, σ ₂ = 1.8%, r ₂ = 7.6% (new prognosis) is illustrated in FIG. 5.

If, at this minimum of the ratio sum, the corresponding ratio quotient is less than one, X ₁ thus lies outside the forecast conflict zone, then the minimum point of the ratio sum is searched for the next point at which it is equal to 1; in the case of two such points X ₁ , X _j are chosen that having the lower ratio.

Subsequently, an entry is sorted into the ratio queue according to the associated ratio quotients and for this purpose the determined optimal expression value X ₁ is noted as a separation value (see block 8.5 in FIG.

The ratio queue is now sorted in ascending order by ratio sums (see block 8.6 in Fig. 3), and after a query (field 8.7), if the ratio queue contains an entry and a query (field 8.8), if the weight In the negative case, the new forecast is eliminated by processing forecast conflict records, whereby the eliminated weight G _n , is determined on the basis of the associated ratios. For this purpose, it is checked whether the new prognosis has the higher relative weight G _n :

G _n X (R _n -I)> G _m

In this case, according to block 8.9 in FIG. 3, the weight of the new forecast is changed

G _n = G _n -G _1n X (R _1n -I)

is reduced, and a prognosis conflict data set is stored with the cut-off value and the two weights G _n , G _n , and ratios R _m , R _m ^c in a memory unit 10 (see Fig. 1), cf. Block 8.10. Then the next entry in the ratio queue is processed.

In the opposite case, the weight G _{n of} the active forecast (block 8.9) is reduced analogously

G ₁₁₁ = G _1n -G _n X (R _n -V) ₁

and the corresponding forecast conflict data record is stored in the memory unit 10 (see FIG. 1) in accordance with block 8.10 with the separation value and the two weights G _m , G _n and ratios R _mr R _m ^c . If the new forecast is completely eliminated (field 8.8 in Fig. 3, output "Yes") or the ratio queue is processed (field 8.7, output "No"), the aggregation of the new forecast information in the aggregation unit 9 (see FIG. Fig. 1) continued.

After contradictory prediction signals have been eliminated in the manner described above, the remaining active forecasts in the aggregation unit 9 (FIG. 1) are aggregated in an appropriate manner, cf. also block 8.11 in FIG. 3. Advantageous aggregation possibilities include the simple method, the center method and the combination method.

- In the simple method, the forecast with the lowest noise factor r is used as the descriptive variable, and from it x ₀ , σ ₀ , L ₀ and r _{0 are} read.

In the middle method, the combination of those predictions P ₁ and Pj, respectively with distribution LN and noise factor r, in which the following two expression values x _u and x _d of the noise factors adjusted medians are respectively minimal and maximal:

The aggregate (total) forecast P ₀ with X ₀ , σ ₀ , G ₀ and r ₀ results from forecasts P ₁ and P _j as follows:

The aggregated forecast value is:

Σ ₀ for the weighted with values G ₁ and G _j average of Oi and O _j holds.

The weight G is the weight sum. The aggregated noise factor R ₀ is calculated as follows, where LN ₀ is the lognormal distribution function for X ₀ and σ _0:

r _o = LN ₀ (X ₁₁ ) -L N ₀ (x _d )

- Combination method combines all active forecasts weighted. In a combination option, the weight sum G of all n active forecasts is calculated.

Alternatively, it is also possible to adjust the weights in order to take less account of individual forecasts with a higher noise factor in the aggregation, for example by dividing G ₁ by Σχ ² .

Now, in the first exemplary combination possibility, the relatively weighted mean x _{0 is} calculated as follows:

The relatively weighted combined standard deviation σ ₀ is then determined as:

σ _o = - Y σ. xσ *

<J; = 1

After an aggregated prognosis information is present, which is supplied according to FIG. 1 to a control unit 12 for the purpose of controlling a process unit 13, and which is preferably also transmitted to the networked forecasting units 2.i and possibly other, downstream prognosis units 2 ', it can be used for a forecasting unit be advantageous to be able to release bound weight from forecasting conflicts for other forecasting purposes if necessary, cf. also block 8.12 in Fig. 3, wherein currently active prognosis signal data sets for weight substitution - in a substitution unit 15 - are used. For substitution, the prediction unit 2 transmits a substitution signal to the substitution unit 15 of the central prognosis system 1, which determines the weight to be released as follows:

First, according to the above-mentioned formula for the ratio calculation, for each active prognosis signal n stored in the prognosis signal memory 4, its ratio (R _n ) relative to the specific prognosis separation value (X ₁ ) of the predicted conflict dataset i to be substituted is determined taking into account the conflict side Φ and entered into a ratio queue for each active forecast. Then each entry of the ratio queue is sorted in ascending order. Thereafter, this ratio queue is processed by calculating the weight releasable by the prognosis substitution as follows:

where Gj _. the weight bound in the prediction conflict i, R ₁ its original ratio of the conflict record, and R _n the ratio of the active forecast from the ratio queue. Insofar as this weight is available in the relevant prognosis signal read from the memory 4, in the weight memory 11 the free (otherwise only available) weight of the original prediction unit i is added and subtracted from the active prognosis signal n.

A new forecast conflict record is stored by figuratively assuming the original prediction unit i the complementary conflict side ^~ Φ and the active forecasting unit occupying the original conflict side with the weight thus determined, now released. In the case of only partial availability of the required weight (G _n <G _free ), the conflict weight G ₁ in the original forecast conflict data set is reduced proportionally (with G _n / G _fre i). Subsequently, the next ratio queue entry for the remaining weight G _{1 is} processed.

As far as the evaluation is concerned, all forecast conflict data sets that have not already been substituted are evaluated at the time of the forecast, cf. the evaluation unit 14 in Fig. 1, by a actual measured value is determined with the aid of a measuring unit 16 or a time-lagged forecast value for the same prognosis size is used, which at most has to be corrected for the time dimension. This is particularly advantageous if, due to a decision or a control pulse influencing the measured value, the measured value can no longer be determined before the prediction time is reached or a destruction test is to be avoided, in which case the last valid aggregated prognosis value (before the influencing control pulse ) can be used.

The total weight bound in the forecast conflict is then assigned in the evaluation process by the evaluation or weighting unit 14 completely to that forecast unit, s. Memory unit 11 in Fig. 1, which was attributed in the elimination of the page right conflict page with respect to the measured value. If the forecast value equals the measured value, the weight is equally divided between the two forecast units.

Furthermore, it can be advantageous, after reaching a prognosis time or even at regular intervals, to additionally attribute a specific weight to each individual prognosis unit 2, independently of the prognosis result, in order to ensure the fate of all prognosis units, albeit with minimal weighting potential in the overall system. A possible variant for this is a percentage increase of the n forecast units as follows:

"N

where GI is the top-up weight, a is a top-up factor (eg, 0.01), and G _{1 is} the current total weight of each of the forecast units. If a = 0, prediction units from the overall system can be completely eliminated. It has proven to be advantageous to set a equal to an average noise factor in order to calibrate the learning effect.

In the hydrological example, it is thus possible to avoid, for example, forecasts of a specialist who specializes in a rare event. The meteorological prediction unit (eg for lightning activity) is no longer considered at all, just because, for example, a long time series without lightning activity arises and forecasts due to the high random factor are evaluated to the detriment of the forecasting unit. In the long run, this can cause the total number of forecasting units to be reduced to one with weight.

From the foregoing, it will be seen that the present technique achieves compression of prediction data such as special coding, with more efficient data transmission and quantified reliable application in various fields of engineering for control systems to financial engineering systems , is possible. It is also conceivable to use the various units shown in Fig. 1, e.g. 5, 8, 9, 14, 15, as a separate units or components, as hardware or firmware, with fixed predetermined processes to realize, as well as a distribution to multiple computers is conceivable.

Claims

Claims :

1. A method for the automated determination of optimized forecasts on the basis of individual forecasts for the control or regulation of operational systems or processes, characterized in that the individual predictions that describe a value for a defined future system or process state and its statistical distribution and the have different weights and noise factors are examined for contradictions, with contradictory individual forecasts eliminated and the remaining individual forecasts are aggregated taking into account the noise factors to an optimized overall prognosis with an associated aggregate noise factor.

2. Method according to claim 1, characterized in that in the investigation for a contradiction in two individual predictions for each prognosis, two ratios (R _m , R _m ^c ) in the form of the inverse probability of a value above or above the noise factor (r) are determined under a certain value (xi).

3. The method according to claim 2, characterized in that each of the two ratios of the later individual forecast is checked to a maximum in relation to the respective complementary ratio of the or each previous active individual forecast, in order to determine a possible contradiction.

4. The method according to any one of claims 1 to 3, characterized in that the weightings of individual forecasts are reduced according to the result of the investigation for contradictions.

5. The method of claim 4 with claim 2 or 3, characterized in that a weight contribution ratio is determined on the basis of the respective ratios.

6. The method according to any one of claims 1 to 5, characterized marked ^¬ characterized in that a prognosis separation value is determined by the fact that for this purpose that expression value (X ₁ ) is determined, in which the sum of the eliminated weightings of two individual forecasts is a maximum.

7. The method according to claim 5 and 6, characterized in that for a prognosis conflict, described by forecast separation value and weight contribution ratio, the sum of the reduced weightings is held.

8. The method according to claim 7, characterized in that in a prognosis conflict the recorded weight of a prognosis (i) is released under weight reduction of another prognosis.

9. The method according to any one of claims 1 to 8, characterized in that the aggregation is based on the single prognosis with the lowest noise factor.

10. The method according to any one of claims 1 to 8, characterized in that in the aggregation those two individual predictions are combined in which expression values of the adjusted with the noise factor medians are minimal or maximum.

11. The method according to any one of claims 1 to 8, characterized in that in aggregation all remaining individual forecasts are combined weighted.

12. The method according to any one of claims 1 to 11, characterized in that an actual measured value or a downstream aggregated overall forecast for the predicted system or process state for evaluation or weight distribution is used on the individual forecast units.

13. System (1) for the automated determination of optimized forecasts for the control or regulation of operational systems or processes (13), characterized by operating in parallel forecasting units (2.i), the output of individual forecasts, a value for a defined describe future system or process state and its statistical distribution, and which have different weightings and noise factors, are set up, connected computer means (3) having an elimination unit (8) for eliminating contradictory individual forecasts and an aggregation unit (9) for aggregating Taking into account the noise factors, the remaining individual forecasts have an optimized overall prognosis with an associated aggregated noise factor.

14. System according to claim 13, characterized in that the elimination unit (8) and the aggregation unit (9) is preceded by a prediction signal processing unit (5).

15. System according to claim 13 or 14, characterized by a separate memory unit (4) for the prediction signal data sets representing the individual forecasts.

16. System according to any one of claims 13 to 15, characterized in that the elimination unit (8) is a memory unit (10) connected to store forecast conflict data records.

17. System according to any one of claims 13 to 16, characterized by a evaluation unit (14) for evaluating the prognosis conflict data sets and, where appropriate, redistribution of the weightings of the forecasting units on the basis of the aggregate total forecast and / or one of a measuring unit (16) supplied measured value.

18. System according to any one of claims 13 to 17, characterized by a substitution unit (15) for allocating weight released from a prognosis conflict to a forecasting unit.