Computer implemented method for predicting energy usage for a route

Abstract

To provide a prediction method of energy used in a route executed by a computer.SOLUTION: A prediction method of the energy includes the steps of: inputting map data of roads included in K trips within a geographic area and a predictor of energy utilization along the roads; inputting energy consumption data representing total energy usage between a start point and an end point in each trip; dividing each road in the map data into segments for all trips and grouping the segments into N clusters which are defined according to the range of at least one predictor; assuming that the known total energy usage of the trip as equivalent to a sum of the known length indicators for each segment in the trip multiplied by the cluster weight, assigning the weight applied to the cluster for each segment as the energy utilization; and storing a segment ID along with the energy usage indication so as to predict route energy used within geographic areas including one or more segments.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention is concerned with predicting the impact of individual components on aggregate statistics (statistics that are sums of other random variables). In particular, it may be concerned with predicting the impact of each segment of a trip on the total energy consumption of an electric vehicle. The present invention may further be concerned with energy-efficient route optimization by applying route optimization algorithms after estimating the energy consumption for each segment of a trip, such as a road segment. [Background technology]

[0002] One of the biggest barriers to large-scale adoption of electric vehicles (EVs) is range anxiety. Despite recent improvements in battery technology and an increasingly wide distribution of charging stations, many consumers worry that miscalculation of their remaining driving range could leave them stranded with a dead battery.

[0003] The reason for this uncertainty is that the energy consumption of electric vehicles depends on many factors, and therefore the energy consumed on the same journey or in the same vehicle can be unpredictable. These factors include, for example, driving style (passive or aggressive), vehicle condition, driver mood (e.g., if the driver's mood on a particular day causes his driving style to deviate from his usual driving style), weather conditions, road characteristics, and unexpected traffic events that occur during the journey. This lack of predictability is exacerbated by errors and noise in the measurement of important characteristics such as battery capacity and acceleration.

[0004] There have been many attempts in the prior art to reduce the scope of the effects of anxiety. The most notable energy consumption estimation model relies on time series data linked to each road segment and ground truth values ​​of energy consumption at the link level for each road segment. This is called link-level ground truth, where ground truth is used here to represent empirical values ​​as opposed to derived measurements / values ​​or known values. GPS loggers have been used to link data to road segments with the aim of predicting their energy consumption. However, such data are expensive to acquire and, when acquired, they require lengthy pre-processing before use in modeling.

[0005] Training a machine learning model for accurate prediction of energy consumption in a supervised manner relies on link-level ground truth values ​​for the energy consumption of each road segment, and similarly for other models. Summary of the Invention

[0006] The inventors have realised that it would be advantageous to solve the problem of predicting energy consumption of road segments without link-level ground truth values.

[0007] Thus, the inventors have realized that it would be desirable to provide a modeling technique that estimates the energy consumption of each segment from aggregated trip data, without link-level ground truth values ​​for the energy consumption at midpoints on the trip. More generally, it is desirable to predict aggregate statistics (e.g., in the fields of energy consumption, finance, production control, or anomaly detection) based on the influence of modeled individual components, but without knowing the contribution of the individual components.

[0008] According to an embodiment of a first aspect of the present invention, there is provided a computer-implemented method for predicting energy usage for a route. The terms "road", "vehicle" and "route" are used generically herein, and a vehicle may be a car, truck, bus or motorcycle moving along a road, or any vehicle moving along a line or other route, that uses energy to move.

[0009] The energy may be for a standard or average vehicle, or may be adjusted for a particular vehicle after or during the method. Similarly, the energy may be adjusted for a predicted speed, or may be based on a normalized or average speed.

[0010] In a first phase, a model is built. There are two kinds of inputs: map data and energy consumption data, which may be in parallel or in any order and successively. The method may comprise inputting map data of the roads included in a number of trips, i.e. K trips, in a geographical area, and inputting a predictor of the rate of energy use along the roads. The method may further comprise inputting (vehicle) energy consumption data of the K trips, the energy consumption data indicating the total energy use, i.e. consumption T, between the start point A and the end point B of each trip, measured for example as power in Watts or as rate per kilometer in Wh / km. The latter data may be averaged over the range of the vehicle or standardized, for example set by the particular vehicle. The roads in the map data are represented by a length index λ i (usually fractional, but possibly real-valued, numbers) of segments, which may have different respective length indices. For example, λ i may be the i-th segment length divided by the total length of the trip, i.e., a number in the interval [0,1]. The segments from the trip may be grouped into a number N of clusters, where the clusters may be defined according to the range of at least one predictor, and each cluster has a weight Wj is defined as having

[0011] The method then calculates weights W based on solving one simultaneous equation for each trip. j An algorithm may be used to build a model predicting the total energy consumption of the trip, where each equation equates the known total energy consumption T of the trip to the sum of the known length indicators of each segment in the trip multiplied by the weight of the cluster into which the segments are grouped. For each segment, the method multiplies the predicted weight W applied to the cluster into which the segment is grouped. j Let Y be the predicted rate of energy use. i and the predicted rate of energy usage Y i An indication of the predicted total energy usage or predicted energy usage rate Y i ) along with the segment ID. This allows for prediction of energy usage for routes within said geographical area incorporating one or more segments. The last step may be included in a query phase, which queries the model to retrieve information relating to the energy usage of each segment. Simple addition of this information per segment (possibly scaled up or down according to a particular vehicle and / or predicted speed) will give an idea of ​​the energy usage of the route. The prediction of energy usage may be compatible (not adopted for a particular vehicle), for example, to select a route from potential routes between a start point and an end point based on energy consumption.

[0012] Optional pre-processing may include, prior to inputting the map data, finding the shortest route between start point A and end point B for each trip in the database, identifying the length in the database between start point A and end point B, and including the shortest route data for a trip if its length is within a threshold of the shortest route between start point A and end point B to form the K trips. The threshold may be expressed, for example, as an absolute length or a percentage deviation. The percentage deviation may be 10%, more preferably between 1 and 5%, and most preferably 2%.

[0013] Optional pre-processing may include data cleaning before inputting the map data, and may desirably include smoothing the elevation data on an elevation tile by modifying the elevation of one or more edge points on that tile that are adjacent to a tile of a different elevation, if there is more than one point on that tile, to bring the elevation of the one or more edge points closer to the elevation of the tile of the different elevation, i.e., the pre-processing may linearly interpolate the elevations of intermediate points while taking into account the elevations of points in different elevation tiles.

[0014] The method may be extended to include all segments in the geographical region by dividing the portions of the road in the geographical region that are not included in the trip into unexplored segments, grouping any unexplored segments into clusters according to the level of at least one predictor in the unexplored segment, and using the length indicator and predicted rate of energy use of each unexplored segment to predict energy use also for routes that incorporate one or more of the unexplored segments. The same method is then used for the previous (unexplored) segments in the model to group the segments into clusters once the model is generated. The method may thus cover all roads on the map, even if they are not going to form part of one of the trips used to build the model.

[0015] Any suitable predictor may be used. For example, the predictor may include at least one of the following: gradient in the segment, such as the average positive or negative gradient of the segment, and features affecting traffic in the segment, such as past traffic data, traffic signals, junctions, shops, road type, weather temperature, vehicle speed, and / or wind impact. The gradient, traffic signals, shops, and road type may be derived from map data (effectively any geographic data that may be map data including roads, but not necessarily the same). The remaining data, such as weather temperature and vehicle speed and / or wind impact, may be input with the energy consumption data associated with the trip. For example, the wind direction in the geographic area may be queried from an open source data source before the trip begins, and the ambient temperature may be obtained from the weather data for the day. With regard to speed, the history of the particular driver and the history of past traffic data (if available) may be taken into account, and if no history is available, the method may use a nominal speed value for each road segment provided with the energy consumption data. Other vehicle data, such as battery temperature, may be obtained from the vehicle and / or energy consumption data.

[0016] Clustering is an important part of the method. Preferably, clusters are defined according to the range of one or more predictors resulting in a predictor space with a number of dimensions equal to the number of predictors. For example, two main predictors that may be used together in a two-dimensional predictor space are gradient and number of stores, or gradient and traffic lights. The predictor space may, for example, have one axis for each predictor. Predictors with a large impact on energy use are preferably divided into more cluster divisions than predictors with a small impact on energy use.

[0017] The method includes providing, for each dimension, a split representing a level of a predictor having a smaller impact on energy consumption at a lower split (and / or at a lower position along the axis for that predictor) than a split representing a level of a predictor having a larger impact on energy consumption, and j W j+1This results in a natural and easily understandable grouping into clusters.

[0018] The method is particularly advantageous when the number M of road segments in the geographical region is at least an order of magnitude greater than the number N of clusters and / or the number K of trips.

[0019] Any suitable algorithm may be used to predict the cluster weights and thus build the model. In an example, the algorithm for building the model may be a quadratic optimization algorithm or may be developed by a neural network.

[0020] If the algorithm is a quadratic optimization algorithm, it may be constrained with constraints on the weights, where the weights are constrained according to the cluster they are assigned to (with more weights for higher clusters) and thus ordered according to the impact each predictor has on the energy consumption rate.

[0021] Even for the same trip from A to B, the energy consumption data may have different values ​​according to the driver's driving behavior / mood, the traffic conditions of the day, the weather, etc. Thus, the same trip may be associated with different values ​​and the energy consumption rate is not averaged over the trip (naturally the predictors will take different values). The energy consumption rate (per kilometer, e.g., and therefore over distance) may be obtained from the energy consumption data or used instead of this value (if the length measure is given as a fraction of the total length).

[0022] A computer implemented method for route optimization may include using predicted energy usage of possible routes within a geographical area between a start point and an end point, calculated according to any of the above definitions, and showing a user the relative energy usage of the possible routes, for example by showing the lowest energy route. Suggested routes may be provided in a navigation aid for the user.

[0023] According to an embodiment of a second aspect of the present invention there is provided a computer program comprising instructions which, when executed by a computer, cause the computer to carry out the method above.

[0024] The program may be executed locally or in the cloud to provide an energy usage forecasting methodology on a local device.

[0025] According to an embodiment of the third aspect of the invention, there is provided a computer (data processing apparatus) having a processor and memory configured to execute the above method, which may be linked to a display (e.g. displaying input fields and results) and to an input device (for a developer to input data and parameters to build a model of energy usage and / or for an end user to indicate the start and end of a route).

[0026] A corresponding computer system may include the above-mentioned computer, display, and input devices, as well as any other necessary components, such as a network interface.

[0027] Finally, in a generalized manner, there is provided a computer implemented method for estimating individual components contributing to an aggregated statistic, comprising inputting data for K trials, inputting a predictor of the overall proportion of contribution to the aggregated statistic, inputting contribution data for the K trials, the contribution data indicating a total contribution T between the start and end of each trial, and inputting each of the K trials as a (fractional) length index λ.i into time or space segments, and grouping the segments from the K trials into N clusters defined according to the range of at least one predictor, each cluster having a weight W j and weights W based on solving one simultaneous equation for each trial. j The objective of the present invention is to use an algorithm to build a model to predict the contribution of a trial, where each equation equates the known total contribution T of a trial to the sum of the known length indices of each segment in the trial multiplied by the weight of the cluster into which the segments are grouped, and for each segment, the predicted weight W that has been applied to the cluster into which the segment is grouped. j Let Y be the predicted contribution ratio. i and the corresponding cluster ID or predicted contribution fraction Y, to allow prediction of the contribution for an aggregated statistic incorporating one or more segments. i and storing the segment ID along with an indication of the segment ID.

[0028] An apparatus (computer or computer system) or computer program according to preferred embodiments of the present invention may have any combination of method aspects. Similarly, features of energy usage aspects may be applied to generalized aspects, except where clearly incompatible. Methods or computer programs according to further embodiments may be described as being implemented by a computer, in that they require processing and memory capabilities.

[0029] An apparatus according to the preferred embodiments may be described as configured or arranged to perform certain functions, or simply as performing. This configuration or arrangement may be through the use of hardware or middleware or any other suitable system. In the preferred embodiments, the configuration or arrangement is through software.

[0030] The invention may be implemented in digital electronic circuitry, or in computer hardware, firmware, or software, or in combinations of them. The invention may also be implemented as a computer program or computer program product, i.e. a computer program comprising instructions tangibly embodied in a non-transitory information carrier, for example a machine-readable storage device, or in a propagated signal, for execution by or to control the operation of one or more hardware modules.

[0031] The computer program may take the form of a stand-alone program, a computer program portion, or more than one computer program, and may be written in any type of programming language, including compiled or interpreted languages, and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a data processing environment. A computer program may be deployed to be executed in one module or in multiple modules at one site or distributed across multiple sites and interconnected by a communication network.

[0032] The method steps of the present invention may be performed by one or more programmable processors executing computer programs to perform the functions of the present invention by operating on input data and generating output. Apparatus of the present invention may be implemented as programmed hardware or as special purpose logic circuitry including, for example, a Field Programmable Gate Array (FPGA) or an Application-Specific Integrated Circuit (ASIC).

[0033] Processors suitable for the execution of a computer program include, by way of example, both general purpose and special purpose microprocessors, and any one or more processors of any kind of digital computer. Generally, a processor receives instructions and data from a read-only memory or a random access memory or both. The essential elements of a computer are a processor for executing instructions coupled to one or more memory devices for storing instructions and data.

[0034] The present invention is described with respect to specific embodiments. Other embodiments are within the scope of the following claims. For example, the steps of the invention may be performed in a different order and still achieve desirable results. Multiple test script versions may be compiled and invoked as a unit without using object-oriented programming techniques. For example, the elements of a script object may be organized in a structured database or file system, and the operations described as being performed by the script object may be performed by a test control program.

[0035] Elements of the invention may be described using "processors," "input devices," etc. As will be apparent to one of skill in the art, such functional terms and their equivalents refer to parts of a system that are spatially separate but combine to provide a defined function. Similarly, the same physical part of a system may provide two or more of the defined functions.

[0036] For example, any separately defined means may be implemented using the same memory and / or processor, where appropriate.

[0037] Preferred features of the present invention will now be described, by way of example only, with reference to the accompanying drawings, in which: [Brief description of the drawings]

[0038] [Figure 1](a), (b) and (c) are overall diagrams of combining the prior art with the new method. [Diagram 2] 1 is a flowchart of a method in a general embodiment. [Diagram 3] FIG. 2 is a diagram of a database used in modeling energy use. [Figure 4] FIG. 1 illustrates the weight adjustments used to approximate the true value from the predictions made by the model. [Diagram 5] FIG. 1 is a process flow diagram illustrating the underlying idea. [Figure 6] Here is an example of a map showing a trip from A to B. [Figure 7] FIG. 1 shows a diagram of a trip divided into segments and segment bins (clusters). [Figure 8] 1 is a bar chart showing the significance of the range of predictions. [Figure 9] This is a table of constraint ranges in a one-dimensional predictor space. [Figure 10] This is a table of constraint ranges in a two-dimensional predictor space. [Figure 11] 1 is a table of segments and the predictors applied to them. [Figure 12] 11 shows the segments of FIG. 10 assigned to clusters (bins). [Figure 13] We show the scaling of the average temperature gradient distribution in photovoltaic applications during clustering. [Figure 14] 1 is an overview of the data used during modeling and during routing. [Figure 15] FIG. 2 is a diagram of suitable hardware for implementation of an embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0039] [Cutting-edge technology] A large body of literature is devoted to the problem of energy-efficient route estimation for electric vehicles. For the prior art, accurate energy consumption estimation requires obtaining vehicle and driver information along with link-level ground truth values ​​for energy consumption at intermediate locations during the trip. This is because over-aggregation of information causes loss of data variance, making it difficult to estimate the contribution of each individual constituent segment to the energy consumption in a trip. Furthermore, it is widely known that vehicle speed and acceleration greatly affect energy consumption, since the most accurate prior art modeling attempts to infer the driver's speed profile by relying on time series data linked to each road segment.

[0040] Sensors linking data to each road segment include GPS loggers with frequencies from 1 Hz up to 10 Hz used to identify latitude / longitude coordinates and instantaneous speed. Acceleration estimates may then be obtained using finite differences between speed values ​​divided by the distance traveled. Having precision in GPS coordinates results in higher accuracy in estimating speed / acceleration and thus energy consumption. Furthermore, sensors for battery condition, battery voltage and degradation rate help determine anomalies that may significantly change energy consumption during the trip. Last but not least, on-the-fly information on traffic conditions and weather (temperature, wind, precipitation) allows driver behavior estimation (speed profile) and auxiliary consumption estimation to be made, which is important since they can have a large impact on energy consumption.

[0041] Many new techniques exist for estimating the energy consumption of electric vehicles (e.g., cars, trucks, vans, campers, or boats or trains not moving on roads or other electric-powered or hybrid vehicles). Such techniques include regressors such as multi linear regression, support vector regressors, gradient boosting, etc., along with Markov chains, physical models, combinations of neural networks and physical models, reinforcement learning, and many others.

[0042] To train machine learning models to predict energy consumption for each road segment in a supervised manner, it has been considered essential to obtain ground truth values ​​for energy consumption at intermediate locations of the trip along with vehicle and driver data, but such detailed information can be difficult or expensive to obtain, or, if obtained, requires lengthy pre-processing.

[0043] Figure 1(a) shows a curve in a road. The points along the curve represent data provided by a GPS logger that can measure position and instantaneous speed. In this prior art example, intermediate ground truth values ​​for energy consumption (link level ground truth, i.e. values ​​for the links or segments between each recorded point) make it easier to infer the energy consumption for that road link or segment.

[0044] An embodiment of the invention does not have GPS information at intermediate positions. This scenario is depicted in FIG. 1(b). Thus, the route that the electric vehicle (EV) will follow is uncertain. Furthermore, only the total Energy Consumption (EC) is known for each trip. FIG. 1(b) illustrates different routes connecting the beginning and end of a road curve. To combat this issue, an embodiment of the invention does not consider routes of mile length that differ from the shortest route connecting the start and end points by more than a certain percentage, e.g., 1-5%, preferably 2%. That is, data from such routes is not included in the method. This is because such routes are unlikely to follow the shortest route that would be considered in the method as the EV's route.

[0045] FIG. 1(c) again shows the points added in the method to partition road segments. Open source route optimization software such as OSRM (Open Source Routing Machine) or GraphHopper may be used to partition the trip into road segments (a map may be viewed as a graph, road trips may be viewed as paths linking the nodes of the graph, and road segments are the links between adjacent nodes). From the aggregated driving information (e.g., average speed, no acceleration information), it is much more difficult and challenging to infer the energy consumption of each road segment without link-level ground truth values ​​of energy. This is because over-aggregation of information causes loss of data variance, making it difficult to estimate the contribution of each individual constituent segment to the energy consumption in the trip. For example, with only two road segments with ECRs of 200Wh / km and 250Wh / km respectively, the average ECR is 225Wh / km, which is not a bad estimate for each road segment, and the amount of information lost is only small. In contrast, with 20 road segments with widely varying road gradients (both positive and negative), the ECR will also vary widely (e.g., 200Wh / km, 70Wh / km, -90Wh / km, etc.). The average ECR will be 60Wh / km, which is a poor estimate for each road segment due to over-aggregation. Embodiments of the invention overcome this problem (and similar general problems when modeling aggregated data), as described in more detail below.

[0046] Of course, standard energy estimates may be modeled for a specific vehicle and take into account specific driving conditions. In the modeling phase, manufacturer data on energy consumption may include one type of vehicle and section speeds. Although the route search engine expects to compare different driving routes, it does not need to rely on such data (speed, wind, driving style, etc.). The most energy efficient route does not change due to these predictors, for example, if there is more wind or the speed is different. The most energy efficient route should be the same in all situations.

[0047] In contrast, to determine the EV distance, more data is needed. The method can use two different models. The first model uses only the gradient, traffic data, shops, etc. to generate a model for the route search engine to find the most energy efficient route. The second model is used for accurate ECR estimation of the most energy efficient route.

[0048] After finding the most energy efficient route using the first model, two questions arise: A) How much energy will it use depending on driving style, weather, speed, etc.? To perform accurate estimation, more predictors (speed, temperature, wind, etc.) are used to build a second model from the manufacturer's data. The predictors are quantized or clustered to solve simultaneous equations to infer the values ​​of each cluster. B) Is there enough energy to perform this route? If not, the EV must reroute to the nearest charging station. That is, the method may find the most energy-efficient route to the nearest charging point and then reroute again after charging.

[0049] FIG. 2 is a flow chart illustrating a general embodiment of the present invention.

[0050] S2 includes inputting map data for the roads included in the K trips within the geographic region and inputting predictors of rates of energy use along the roads. Some predictors of energy use may be geographic or map data, such as slope.

[0051] S4 includes inputting energy consumption data for K trips, the energy consumption data indicating the total energy usage T between the start point A and the end point B of each trip. The energy consumption data is provided in a vehicle data set, for example, with empirical data for a particular vehicle trip giving the start and end points of the trip and mileage which is matched against map data to derive the roads used, and optionally average speed (other predictors) and other relevant factors.

[0052] In step S6, each of the roads in the map data for every trip is assigned a length index λ i It is divided into the following segments.

[0053] In step S8, the segments from the trip are grouped into a number N of clusters. The clusters are defined according to the range of at least one predictor, and each cluster has a weight W, which is to be determined (using an algorithm). j has.

[0054] In step S10, the algorithm calculates the weights W by solving one simultaneous equation for each trip. j We build a model to predict,T,=,T,. Each equation equates the known total energy consumption of a trip,T,to the sum of the known length indices of each segment in the trip multiplied by the weight of the cluster into which the segments are grouped.

[0055] In step S12, for each segment, the predicted weight W applied to the cluster into which the segment is grouped is calculated. j is the predicted rate of energy use Y iand the method further comprises: assigning a predicted rate of energy usage Y to the route within a geographic region incorporating one or more segments such that a prediction of energy usage for the route within a geographic region incorporating one or more segments is enabled. i The segment ID is stored along with the instruction.

[0056] Thus, querying the model allows the energy usage of a route to be determined. The querying may be remote.

[0057] Important features of the above method may include: - Prediction of energy consumption for each road segment from aggregated driving statistics without relying on link-level ground truth values ​​for energy consumption. Furthermore, there is no need for intermediate (link-level) driving data such as speed, acceleration, etc. Clustering the predictor space (which is a multidimensional space in which all possible values ​​of the predictors of rate of energy usage can be plotted) into clusters / bins. ● Use of algorithmic techniques such as constrained optimization or projected gradient descent to obtain natural weights in the model. This leads to interpretability of the weights after training, which represent the energy consumption rate of each road segment. • A method to recover the contribution of each component in the aggregated statistics. No reliance on past time series data of the same vehicle for estimation purposes. Hence, embodiments may completely avoid the use of time series data to map journeys to segments.

[0058] [Calculation explanation of the underlying idea] The embodiment provides a method for estimating the contribution of each component to an aggregate statistic T. The aggregate statistic T is a function of the contribution of a random variable Y from other random variables. i It is defined as the sum of:

number

[0059] Correspondingly, to predict the contribution over time to an aggregate statistic, T is the total contribution known for any trial in the data set, and Y i is the contribution of segment i over time, M is the total number of possible segments, K is the number of trials in the dataset, and λ i is the (known) length of time indicator for the segment.

[0060] The goal is to solve the following simultaneous equations: (k) and λ i (k) Given the total ground truth value for Y i To guess:

number

[0061] The difficulty in solving the latter system in Eq. 1 is that it is highly uncertain and any data-driven method is likely to overfit. This section uses a clustering of the predictor space to reduce the system of equations in Eq. 1 to a tractable level. i We introduce a method for modeling the energy function (Energy Function) in terms of several variables in the presence of constraints on those variables. Furthermore, constrained optimization (optimizing an energy function with respect to several variables in the presence of constraints on those variables) may be used in areas where there is a need to obtain a weighted model that represents the physical quantities.

[0062] The method may be used to estimate the energy consumption for each road segment without link-level ground truth values ​​for energy consumption or vehicle / driving data. The goal in this exercise is to estimate the energy consumption for each road segment Y against the total energy consumption T of each trip. i The difficulty in doing so is the nature of the data and, in general, the unknown Y i There is much less data available than the number of

[0063] The main idea is to reduce the number of unknowns and make the system of equations in Eq. 1 more tractable by solving for the (path) segment Y i (i=1, ,M) can be classified into bins W j The solution is to group the road segments into groups of similar segments called (j=1,...,N). One can then use techniques such as constrained optimization to infer the bin values ​​to have natural weights in the model, where the model weights represent the energy consumption rates for each road segment.

[0064] As mentioned above, the energy consumption rate Y iis influenced by various factors, at least some of which may be publicly available for each road segment. For example, available geographic data may include 30-meter mesh elevation data along with store locations, and traffic signals in a map may be input to the model from public sources. For weather, hourly ambient temperature, wind and rainfall information may be used. Finally, the car manufacturer and / or other sources may provide road trip data including trip start and end points, trip distance, average speed, auxiliary consumption, idling time, and total ground truth values ​​for total energy consumption (note that ground truth values ​​for energy consumption at intermediate locations are not known, referred to as link-level ground truth). A data diagram is given in FIG. 3 and shows the database used in the modeling, which may include any of the geographic, weather and aggregated car manufacturer data listed above.

[0065] Predictors may be inferred from existing data sources. For example, road junctions may be used as predictors, which may be implemented, for example, by adding a Boolean value to each node in the road graph (a segment is a road section between two nodes) describing whether the node is a junction or not. For example, if at least one of the road segment nodes is a junction, place a value of 1 (indicating the presence of a junction) on that road segment, and 0 otherwise. Alternatively, for each segment, the method may determine the total number of adjacent segments, which is simple to implement, for example, using OSRM or Graph Hopper. The predictor may then be defined by counting how many adjacent road segments a road segment has (a segment with more than two adjacent segments indicates the presence of a junction). The more congested the segment is, the higher the ECR is expected to be (due to idling or breakdowns).

[0066] One objective is to estimate the energy consumption for each road segment in a map, e.g., a map of Japan, possibly even for road segments that are not present in any of the trips in the dataset. This can be achieved by clustering segments that are not included in any of the trips according to a predictor and giving them a weight from the cluster to which they are assigned. The estimated values ​​are then run through a shortest route finding algorithm (e.g., Dijkstra's algorithm or ALPHA) to find the most energy-efficient route between two points in the map. * ) may be used.

[0067] This method introduces a robust way to model each road segment and infer their respective energy consumption. Clustering of predictors in a multidimensional predictor space may be used to make the problem tractable, and constrained optimization may be used so that the energy consumption weights of the model have natural values. This method may be used to address any problem that requires modeling and estimation of individual components that contribute to an aggregated statistic. In many real-life situations, there is a need to perform time series forecasting using aggregated information. The goal is to infer values ​​of a response variable at intermediate time steps (hourly, daily, etc.) using information aggregated over a larger time frame (e.g., daily, monthly, or yearly). This frequently occurs in areas such as finance, production management, and anomaly detection.

[0068] [Detailed calculations] As mentioned above, the purpose of this method is to obtain aggregated statistics.

number

number

number

[0069] conditions

number

number

number

number

number

[0070] The goal is to know only the total ground truth value for T and to find the i To recover the influence that each Y i Therefore, the goal is to find the value of T for all k = 1, 2,...,K. (k) and λ i (k) Considering the true value of Y i Simultaneous equations for

number

[0071] T (k) Let Y denote the total ground truth value for the energy consumption rate for the kth trip in the data, and Y i Let λ represent the energy consumption rate of a road segment in a map. The Energy Consumption Rate (ECR) is defined as the energy consumption (difference in battery capacity at the start and end) divided by the total length of the trip. There may be billions (or more) of road segments (M) in the map to infer, but we only have information about a few thousand in, say, a trip K. When modeling each road trip, λ i (k) is zero if the ith road segment in the map is present in the trip. Therefore, the system of equations is also extremely sparse. Thus, K is on the order of thousands, and M (the total number of distinct road segments in the map) is on the order of billions (or more), and the system of equations is largely underdetermined. To indicate that M is orders of magnitude larger than K, we write M>>K.

[0072] To reduce the number of unknowns M and make the inference tractable, we segment Y into W clusters. i A grouping of M is provided using a clustering-binning technique, where each cluster contains N segments, M>>N, and weights W1,...,W N By Y i represents the N clusters into which i (k) Let denote the cumulative fraction that each cluster type appears in the kth expression, i.e.,

number

[0073] This loop repeats for the kth cluster with length index d i (k) We simply add the length indices of the segments of each cluster type that are included together in the trip to give {overscore (x)}, which in this case is a fractional length index.

[0074] Then, the simultaneous equations in Equation 2 are

number

[0075] d j (k) Note that is zero if the jth bin does not appear in the kth sample / trip. Note that previously the sum was over all possible components M (M>>N), whereas now the sum is over all possible bins N. The difference is that the unknown weights W j The simultaneous equations in this case are tractable, since the total number of (obviously N, i.e., the same number as the number of bins) can be appropriately selected during clustering-binning to be comparable (or smaller) to the number of points / trials K (N may be selected, for example, close to or smaller than K).

[0076] The goal in this case is to find d j (k) and TrueVal (k) =T (k) In light of this, j Recall that the aim of the present invention is to find the value of . Any optimization algorithm or neural network architecture may be used to solve the above simultaneous equations in Equation 3. Here, the recommended method is to use the wj We use a constrained quadratic optimization algorithm on and the "OSQP" solver with a maximum of 50K iterations to fit the data.

[0077] Using a neural network, training is as follows: During training, the model predicts

number

number

[0078] Figure 4 illustrates the weight adjustments used to get closer to the true value from the predictions made by the model. Constrained optimization in neural networks can be achieved by projected gradient descent techniques, which are similar to standard backpropagation techniques, the only difference being that we project the gradient descent updates onto the feasible region of the constrained space.

[0079] Weight W i After determining the weights (per bin-cluster), every road segment in the map (including those that do not form part of any trip) may be classified into a particular cluster and therefore associated with a weight value. This information is saved in a dictionary {road_ID:bin-cluster} and used to find the most energy-efficient route (Dijkstra algorithm or A *This may be used in a route engine that implements a route optimization algorithm (such as

[0080] A process flow diagram illustrating the underlying idea in one embodiment is given in Figure 5. A database supplies data for optional pre-processing (depending on the input data set), which uses an optimal path finder to find shortest distance trips (S100) and optionally performs data cleaning including, for example, elevation flattening and outlier detection, and keeps only trips with distances close to the shortest path (S110).

[0081] The pre-processed selected trips are fed into a stage of modeling clusters. In this stage, a cluster algorithm assigns the trips to bins / clusters (S120). This allows the equations to be transformed into equations that are relative to the bins (S130). A constrained optimization solves the transformed equations in S140 and the system can then be queried for the best route between different locations in S150.

[0082] [Application of the method to energy-efficient route optimization] This example determines the contribution of each road segment to the total energy consumption of the trip using only aggregated information, without link-level ground truth values ​​for energy consumption (EC) (see Figure 1c), and without driving / vehicle information at midpoints.

[0083] Since determining energy consumption is equivalent to determining the energy consumption rate (ECR=EC / distance) given distance, the goal is to estimate the ECR of each road segment given the total ground truth value of ECR ​​for the trip. The ECR can be written in a standardized / averaged manner and then compensated to account for EV / driver specific characteristics.

[0084] Total ground truth values ​​of energy consumption rate for a road segment as well as estimates of average speed and auxiliary consumption are provided, for example, by the major vehicle manufacturers.

[0085] The exact route taken by the EV is unknown because GPS was not used at the intermediate positions of each trip in the trip database. To handle this uncertainty, the method retains only a small percentage of trips whose mileage reported by the vehicle manufacturer differed by less than 2% from the shortest route connecting the start and end points of each trip. The shortest route connecting the start and end points of each trip was found using shortest distance route finder software in the graph and then compared with the actual mileage. If the difference with the actual mileage was at most 2%, the driver was considered to have followed the shortest route, otherwise the trip was discarded. Note that there is still a 2% uncertainty in the exact route taken by the EV, which becomes the total error of the modeling approach. Of course, if there is a GPS that measures the exact route, the performance will be improved by 2% and there is no need to remove trips. Further cleaning steps are then performed, such as elevation flattening, removing corrupted data, and outlier detection.

[0086] FIG. 6 shows a map with trips from A to B displayed. The method first filters the trips and keeps those trips with mileages close to the shortest route of the trip from A to B. The method further decomposes the filtered trips into their constituent route segments using available geographic data before clustering them. For each trip, the method may use the information shown, i.e., the energy consumption for the whole trip (but not for the waypoints) and, for each road segment, the gradient, traffic signals, nearby shops, and any other useful geographic data. The energy consumption rate of each road segment may be considered as a function of such predictors.

[0087] Each road segment is the portion of a trip between two points as shown. To calculate the slope, geographic data (e.g., a 30m mesh of elevation data on tiles) may be obtained from any publicly available source (e.g., JAXA's https: / / www.eorc.jaxa.jp / ALOS / en / aw3d30 / index.htm) with additional elevation flattening.

[0088] Here, a route segment is a small portion of the road graph that belongs to the current trip. It is possible to extract (by using an open source route search engine like Graph Hopper) the waypoints in the road graph for the shortest trip between two points A and B. The same procedure may alternatively use a GPS logger.

[0089] The energy consumption rate (ECR) of each road segment is Y i and the aggregated statistic T is the total ground truth ECR for the entire trip. Therefore,

number

number

number

[0090] To make the system of equations more tractable, the method is i Reduce the number of unknowns by grouping x, y, y, into bins or clusters. Any clustering technique for the road segments that minimizes information loss may be used to make the latter system in Equation 4 tractable.

[0091] After clustering, for each Y i (i=1, ,M) is the number of nodes in a particular bin-cluster W j (j=1,...,N). These clusters may represent groupings of edges with properties that result in similar energy consumption rates. For example, one cluster may include road segments with gradients in the range [0,0.05), number of traffic lights in the range [2,5), number of stores in the range [5,10), etc. These clusters effectively define types of road segments and can be imagined as volumes in a multidimensional predictor space. They may be specified in advance by a clustering algorithm, for example using the methods in the co-pending and co-owned applications mentioned above, or by domain experts, depending on the use case. Any quantization-clustering approach results in loss of information, so it is important to choose a grouping with minimal information loss.

[0092] Then, for each trip in the dataset, the cumulative distance of each segment cluster is kept in a dictionary, and a table is generated with columns that represent the corresponding distance divided by the total length of the trip. For example, Figure 7 shows a diagram of a trip of length 22 divided into segments (edges) and assigned to clusters as "bins" (bins). The diagram also shows the length of each segment as L iIn this case, the following dictionary records the cumulative segment length index for each of clusters 1 through N for the associated trip: {1:0,2:2 / 22,3:5 / 22,4:6 / 22,5:9 / 22,6:0,7:0,...,N:0} For example, here clusters 1, 6 and 7 have no segments. Cluster 2 has a total length of 2 (out of a total length of 22). Cluster 3 has a total length of 5 and consists of 2 separate segments (out of a total length of 22).

[0093] From the dictionary, the following formula is obtained to take into account the length of the segments in each cluster and the weights for each cluster:

number

number

[0094] This trip is then added to a table with columns of clusters from 1 to a maximum of N using the dictionary by creating a row in the table and adding the value of each cluster as the corresponding dictionary value.

[0095] The method proceeds as described above for every road trip k=1,...,K available in the dataset. The table will then have K rows and (N+1) columns, with one column being the final sum per trip for the total ground truth ECR, and the remaining N columns representing the different types of bins.

[0096] In one arrangement, these values ​​may be inputs to a machine learning model to attempt to predict the energy consumption rate (and therefore energy consumption) for each type or cluster of road segments. For example, when training a one-layer neural network, the model may predict

number

number

[0097] Weight W i After finding (ECR of each road segment-bin), a lookup table may be generated, which may be passed to, for example, a route engine, which may calculate the energy consumption rate for a given route and implement, for example, a route optimization algorithm for energy-efficient route optimization.

[0098] Alternatively, you can use the wp_cvxpy library (https: / / www.cvxyp.org / ) from python. j A quadratic optimization algorithm with constraints on and an "OSQP" solver with a maximum of 50K iterations for fitting the data may be used. The constraints may be chosen so that the resulting weights (after fitting) reflect the energy consumption of each road segment and are as natural as possible. Furthermore, an ordering of the bins W[i]≦W[i+1] may be introduced according to the influence each predictor has on the problem (e.g., if a bin has a gradient of 0.2 and another bin has a gradient of 0.1, all other predictors being the same, the latter bin will be associated with a lower ECR and therefore a lower W[i]). This ordering may take into account the importance of the predictors and the natural ordering within each predictor.

[0099] The goal of the ordering is for the model weights to represent the energy consumption of the road segments. This increases the interpretability of the model and makes the results more natural. The influence of each predictor (and therefore the ordering) can be determined automatically (and therefore is evident from the data) by using a random forest algorithm. For example, roughly 2000 trees may be trained on the data (X,Y), where X is the aggregated predictor across the trip (e.g., the average slope of the trip, the total number of stores divided by the distance, the total number of traffic signals divided by the distance, etc.), and Y is the ground truth value of the energy consumption of such a trip. The importance of each feature can be measured by plotting a feature importance diagram after training. Features that are more important should be split into more clusters when clustering the values, and those split values ​​may be determined as described in the co-pending and co-owned application mentioned above.

[0100] As an example, the gradients can be divided into 8 bins, each with a given gradient range: B1=[-∞,-0.25),B2=[-0.25,-0.12), B3=[-0.12,-0.05),B4=[-0.05,0), B5=[0,+0.05),B6=[0.05,0.12), B7=[0.12,0.25),B8=[0.25,∞) , where the ordering means that the weights for binning 1 are smallest and W1≦W2≦W3≦···≦W8. i is the "weight", i.e., bin B i The effect of ordering is to make the model results interpretable and natural: a lower gradient corresponds to a lower energy consumption (W i ).

[0101] The number of stores may be divided into four bins: [0,5), [5,10), [10,20), [20,+∞). Again, the ordering of the weights means that the model should follow the rule that a smaller number of stores should be associated with lower energy consumption.

[0102] The total number of bins (8 for gradients and 4 for stores) may be selected according to feature importance after training the random forest. As gradients have higher importance, they were divided into more bins. The total number of bins is a hyperparameter that needs to be tuned by considering the number of data available to the user, and usually requires cross-validation.

[0103] An example feature importance diagram generated from real data can be seen in Figure 8. To increase importance, the predictors are rounded average road type, wind impact per link (segment), number of shops on the link, number of traffic lights on the link, temperature, Aerodynamic Factor (AF), elevation change multiplied by distance, and average elevation change over distance. AF is

number

number

[0104] For each bin, the constraints are imposed in the form of a matrix. For example, if we only use gradients as predictors, the constraints are summarized in a one-dimensional array, as shown in Figure 9.

[0105] The weights reflect the energy consumption rate of the road segment according to the gradient at the corresponding row position. The constraints take the form of intervals by specifying lower and upper bounds for each bin. For example, the energy consumption should be between -600 and -300 for gradients smaller than -0.25. With two predictors, e.g., gradient and number of stops, the constraints take the form of a two-dimensional matrix, as shown in Fig. 10. In a more general sense, this is a hypercube with a suitable ordering. Fig. 10 illustrates this ordering well. Each bin (in the two-dimensional matrix of Fig. 10) is associated with a range of ECR ​​values ​​with a mean value (midpoint). The ordering is implicit here. The bins can be ordered in ascending or descending order according to the average ECR.

[0106] In general, with N predictors, the constraints become an N-dimensional array (hypercube) whose values ​​can be set via a 'for' loop. Note that the upper and lower weight values ​​are problem specific and may be adjusted by the user (as hyperparameters) or by a domain expert. A grid search may also be applied that greedily adjusts the ranges one at a time. For example, the user may specify a step along with the lower and upper bounds of the search space, in which case the grid search will try all possible combinations. In one dimension this may involve specifying a lower bound (e.g. -400), an upper bound (e.g. 400), and a step size (e.g. 100). In that case the grid search will look at all combinations [-400,-400], [-300,-200],..., [300,400]. There are a total of 4 x 3 = 12 combinations (X1 can be in four different intervals: [-200,-100], [-100,0], [0,100], [100,200], and X2 can be in three intervals: [10,20], [20,30], [30,40]). The combinations are: X1 in [-200,-100], X2 in [10,20] X1 in [-200,-100], X2 in [20,30] X1 in [-200,-100], X2 in [30,40] X1 in [-100,0], X2 in [10,20] X1 in [-100,0], X2 in [20,30] X1 in [-100,0], X2 in [30,40] X1 in [0,100], X2 in [10,20] X1 in [0,100], X2 in [20,30] X1 in [0,100], X2 in [30,40] X1 in [100,200], X2 in [10,20] X1 in [100,200], X2 in [20,30] X1 in [100,200], X2 in [30,40] Those skilled in the art will appreciate that higher dimensional lattice searches work similarly.

[0107] The clusters are generated by dividing the N-dimensional space into clusters, each encompassing a range of a respective predictor. In one example where one predictor is used, the number N is equal to the number of clusters. In an example where two predictors are used, first N1 intervals are given, then N2 intervals, resulting in a total of N1×N2 clusters, e.g., as explained above.

[0108] By imposing constraints on the model weights, a more natural fit is obtained. Note that in previous modeling techniques, the weights of the neural network have a physical interpretation: they represent the energy consumption rates of the bin-clusters of each road link. In this way, it is easier to understand the predictions made by the neural network and to interpret the weights after training. Understanding neural network predictions is much more important in other fields such as medicine or autonomous driving.

[0109] A route search engine may then be used to generate a route. The route search engine may use two different models. A first model may only use, for example, gradients, traffic data, shops, etc. to build a model for the route search engine to find the most energy-efficient route. A second model may be used for cumulative ECR estimation of the most energy-efficient route.

[0110] After finding the most energy-efficient route using the first model, two questions arise: how much energy will the route use depending on driving style, weather, speed, etc., and does the vehicle currently have enough energy to perform this route? To perform an accurate estimation, the method may use more predictors (speed, temperature, wind, etc.) to build a second model, e.g., from manufacturer data. The method may quantize or cluster the predictors and solve the simultaneous equations (above) to infer the value of each cluster. If the vehicle does not have enough energy, it must be rerouted to the nearest charging station. That is, the method may find the most energy-efficient route to the nearest charging point and then reroute again after charging.

[0111] [Elevation flattening] Some arrangements use elevation flattening. Elevation flattening may work as follows: Suppose the elevations obtained from a 30m mesh tile of road elevations included in a trip have the following form: [5,5,5,10,20,20,20,20,0], where the numbers represent the elevation in meters above sea level. The first constant 5,5,5 arises because the road trip points belong to the same elevation tile (a tile consists of points at the same elevation). The algorithm "flattens" the elevations as follows: The goal is to change the second and third elevations in the list of elevations to obtain a smoother fiber from the first point A with an elevation of 5m to the fourth point B with an elevation of 10m (similarly flattening from the fifth point with an elevation of 20m to the last ninth point with an elevation of 0m). Consider the points [A,X,Y,B] with elevations "5,5,5,10" respectively, which make up the first four points in the list above. Let E(X) denote the elevation of point X, then, [E(A)=5,E(X)=5,E(Y)=5,E(B)=10]. Changing E(X) and E(Y) will "smooth out" the transition from A to B. So, let E(x)=x and E(Y)=y, and find x and y. Linear interpolation is: (xE(A)) / (AX)=(E(B)-E(A)) / (AB) where (AX) represents the distance between points A and X, and (AB) represents the distance between A and B. These distances are known from the map, so x can be calculated. For example, if (AX)=15m and (AB)=50m, then x=15×((10-5) / 50)+5=6.5m altitude. Similarly, for point Y: (xyE(A)) / (AY)=(E(B)-E(A)) / (AB) and y can be calculated since the distances of (AY) and (AB) are known. If (AY)=35m then y=8.5m. Note that the transition between elevations [5,10,20] does not require flattening since those points belong to different elevation tiles. Finally, the same calculations are performed for the elevation flattening for the first 20m tile and the last 0m tile, i.e., points with elevations [20,20,20,20,0], giving, for example, a flattening of [20,13,7,3,0].

[0112] [Algorithmic parameters for energy usage allocation] Energy consumption is measured in Wh (Watt-hours) and reported in the data set, and since the total length of each trip is also known, it is possible to calculate the energy consumption rate, measured in Wh / km, as energy consumption divided by distance.

[0113] The placement may use the OSQP solver from python's "cvxpy" for sparse quadratic constrained optimization. Default parameters or lasso and ridge regularization may be used. Default parameters (e.g., specifying an upper bound of 50K iterations, specifying an L2 loss, and using a matrix of inequalities (upper and lower bounds) for constrained optimization) work well because the fitted weights have a natural meaning, i.e., they represent the energy consumption rate of a road segment. Since OSQP is used for constrained optimization, a matrix with the minimum and minimum values ​​of each predictor (this is problem-specific and dataset-dependent) may be used to obtain a table like that of FIG. 10 for the two-dimensional case. Thus, for each "binning", upper and lower bounds on energy consumption should be specified. Clusters are formed from the binning. For example, each cell in the table of FIG. 10 may represent a cluster. Alternatively, multiple cells (e.g., the first three groups shown in the top left of this table) may be combined to form clusters, although this can introduce a loss of accuracy and is not always desirable.

[0114] The default parameters for the OSQP solver require the total number of iterations (e.g., upper bound 50K), the loss (e.g., L2, with the user being able to choose L2+ Ridge or L2+ Lasso regularization), and a matrix of inequalities (upper and lower bounds) for constrained optimization. See https: / / osqp.org / docs / examples / least-squares.html.

[0115] In neural networks, it may be difficult to apply constrained optimization with projected gradient descent. Nevertheless, a suitable system may use a forward propagation neural network using 2 to up to 4 hidden layers with 200 hidden units. Of course, one skilled in the art will understand that this design and these parameters will vary across different datasets. For example, with more data, a deeper architecture with more layers and more hidden units may be required to be more accurate. The dataset input (for other algorithms as well) may include approximately 61000 trips after cleaning, and the hyperparameters may use an Adam optimizer with a learning rate of 0.001, again according to the dataset (of course, learning rates of 0.0001, 0.00001, etc. may be used, and convergence is the goal). The total iterations may be 20 epochs. Training may be performed using the PyTorch deep learning framework.

[0116] The hardware used may be a GPU, for example a 1080Ti GeForce 11GB.

[0117] As an example, each trip is split into road segments after querying the graph hopper engine, and the values ​​of each predictor are stored in a table. Each column of the table in Figure 11 corresponds to each predictor, and each row corresponds to each road segment present in the road trip. The parameters fromID and toID correspond to points in the graph (each ID is associated with a (lon,lat) i.e., latitude and longitude pair). Here, segments 0 to 3 are displayed with their start and end location IDs, along with traffic signals in the segment (SignalOverDist), shops in the segment (ShopOverDist), AFs in the segment (AFOverDist) (averaged from the data for the entire trip), road type (TypeRoad) such as unknown, residential, unclassified, tertiary, secondary, primary, arterial or highway (determined from map data), weather temperature (Temperature) and wind effect (taken from weather data), as well as distance and elevation change times distance (EleChangeTimesDistPerLink) (giving the gradient in the last column). The segments are of different lengths, since the nodes in the graph hopper mainly represent junctions or important areas (e.g., landmarks). If this is of interest, the user can query the graph hopper for a road trip and see the various nodes and where they are located.

[0118] Afterwards, a quantization module is used to classify each line (i.e., each road segment) into a corresponding cluster, as shown in Figure 12. Finally, the OSQP solver fits the data with inequality constraints, as shown in the table in Figure 10. The sum of the contributions of the clusters in a trip should be equal to the energy of the trip, i.e.:

number

number

[0119] [Application of the method to time series data for energy generation estimation] The embodiments may be used to estimate the energy generated from photovoltaic (PV) cells in a data-driven manner from aggregated time series data at finer time intervals. There are many energy theory models in the literature that attempt to accurately estimate PV energy production based on an energy equation.

[0120] Suppose it is desired to determine / predict the energy generated from PV per minute (or per hour) given daily ground truth values ​​of the generated energy. For example, at the end of a day, it may be known that a panel generated XWh (watt-hours), but it remains unknown how much energy the panel will generate per minute on a particular day in the future. An estimate of X / 1440 (since a day has 1440 minutes) is inaccurate since the energy depends not only on panel-specific factors but also on many environmental variables. For example, on cloudy days or at night, the energy generation is expected to be close to zero, but when it is sunny, the generation can be very high. Furthermore, environmental variables change dynamically even during a day, since an initially sunny day can become cloudy, and thus different times of the day will have different energy generation (different generation in watts, and therefore different energy generation rates in watt-hours).

[0121] The embodiment used for EV applications above differs in domain from the embodiment used for PV applications (EV is space, PV is time). The EV-friendly embodiment estimates the energy consumption of a road segment knowing the total energy consumption of the trip, and to do so partitions the predictor with respect to the spatial domain. Conversely, the PV-friendly embodiment estimates the energy consumption with respect to a time interval having ground truth values ​​of the energy consumption at the end of a day, or of the energy consumption aggregated over an extended period (accumulation of data trials). Therefore, it is necessary to partition the predictor with respect to the time domain. In addition to this difference, the switch from space to time domain utilizes the same mathematical formulas of the above embodiment. As usual, the sum can be a rate (over time in this case) or an absolute value.

[0122] The energy generation rate in time t is Y t Yt=EG / t, where EG represents the generated energy in units of Wn. Let T be the total ground truth energy generation rate. For example, if the total energy generated in a day is 2880Wh, then the energy generation rate is 2Wh / min (or equivalently, 120W). This aggregated energy generation rate is not a good estimator for a particular time interval of a day; during sunny days, the generation may be 300W, while during cloudy days, the value may be 20W (these values ​​are assumptions for illustration purposes).

[0123] Therefore, similar to the EV equation above,

number

number

[0124] The division of the time intervals may not be uniform depending on the needs of the application. For example, an embodiment may divide the morning and early afternoon hours into 10 intervals, while the evening may be divided into 5 intervals, the night into 2 intervals, etc. The logic should generate more bins per time interval with a higher amount of variability regarding environmental conditions that affect the energy generation or some other strategy depending on the user's intended application. That is, each interval or segment has a unique (time) length.

[0125] The goal is to find Y given the total ground truth value T for the EGR in a day. i The goal is to estimate the value of λ i is known from environmental data. i may describe the time interval during which the environmental variables apply. For example, if there is a sunny day with a temperature of 24°C for 174 minutes, then we can say λ1=174 / 1440. Here, the first variable describes this environmental situation. The total energy generated, T, is then calculated for the i-th interval of the day, λ i The weighted average of energy generation for all different environmental conditions that apply for Y i become.

[0126] Therefore, similar to EV applications, the goal is to solve the simultaneous equations:

number

[0127] According to the equation used in the embodiment suitable for EVs above, we define the energy generation rate Y as a function of several predictors X1, ,X n is considered as a function of . That is,

number

[0128] Any clustering technique suitable for the environmental data and the solar panel specific data that minimizes information loss may be used to formulate the simultaneous equations in Equation 4 (applied to an embodiment suitable for PV energy generation estimation). For example, the following clustering method (grouping of segments from trials) may be used:

[0129] 1. Construct aggregated predictors from geographic data and solar panel type, e.g.: a) Average sky temperature b) Average ground temperature c) Average PV glass temperature d) Average solar angle e) The part of the day that was sunny f) Any part of the day that was cloudy (or any other condition such as rain or snow) g) Average solar panel area h) Average Electrical Energy Conversion Rate i) Others that depend on immediate data Get the.

[0130] 2. Train a random forest on the aggregated data to predict the total energy consumption rate of a trip.

[0131] 3. Track the feature values ​​used to split the nodes of each tree in the random forest in a list-like data structure. For example: -Average temperature: [32,15,2,8,21,...] -Fraction of the day that was sunny [0.2, 0.05, 0.15, 0.3, 0.5, ] -Fraction of the day that was cloudy [0.1, 0.4, 0.05, 0.2, ]

[0132] 4. Generate a histogram of split values ​​for each predictor.

[0133] 5. Find the peak in the histogram.

[0134] 6. Shifting by the difference in means and scaling by the ratio of standard deviations of the distributions (see, for example, Figure 13):

number

[0135] 7. Quantize according to the scale peak value. Then, the simultaneous equation (Equation 1) is transformed into (Equation 3):

number

[0136] 8. Solve the system, for example using the OSQP solver from the cvxpy library (if you want to impose constraints) or a feedforward neural network (if no constraints are imposed). The architecture depends on the type and size of the dataset and the number of predictors. Regularization techniques such as dropout, ridge (if it makes sense to keep weights small in the problem), or lasso (if you want to impose sparsity) may be used. These regularization techniques may also be applied to constrained optimization.

[0137] Similar to the EV embodiment described in the previous section, solving the transformed system of equations (step 7 above) results in predicted cluster weights, which can be used in PV energy generation forecasting.

[0138] [General application of using aggregated data to infer values ​​of quantities at finer time / space intervals] In general, embodiments may be applied to any problem involving aggregated time / space data, where the goal is to predict the impact of some predictors at finer time / space periods. For example, in a finance-related context, Energy Generated (EG) may be replaced by Revenue, and then Energy Generation Rate (EGR) may be replaced by Revenue Rate (IR).

[0139] We have daily income values ​​and several predictors X1, ,X n (Predictors can be anything that the user thinks or knows has an effect on income.) Let Y be the income rate in time t. t That is, Y t = I / t, where I represents income. Let T be the total ground truth value of the income rate. Then,

number

number

number

number

number

[0140] [summary] In one definition, a computer-implemented method for route optimization is provided, comprising: collecting aggregate route segment data including geographic data and vehicle data including vehicle energy consumption rates for vehicle trips; clustering the trip segments into a number of bins, where the trip segments within each bin share data characteristics that result in similar energy consumption rates; and estimating a vehicle energy consumption rate of at least one individual route segment by comparison with the clustered trip segments.

[0141] The method may bring a modeling approach to robust estimation using clustering to make the inaccessible simultaneous equations tractable and recover the impact of each road segment on the total energy consumption of the trip. Furthermore, using any constrained optimization, the method may provide natural weight values ​​for the model that represent the energy consumption rate of each road segment. This makes the model more interpretable and makes it easier to understand the predictions made.

[0142] This method can eliminate the need to obtain expensive link-level data and further saves time since no pre-processing step is required for time-series data.After predicting the contribution of each road segment to energy consumption, a route optimization algorithm may be applied for energy-efficient route calculation.

[0143] The same method steps may be applied to forecast PV energy generation or in financial forecasting.

[0144] In general, the method estimates the impact of each contribution on an aggregated statistic, whether in the time or spatial domain. The proposed solution may therefore be applied in many fields where it is necessary to recover the impact of individual variables on an aggregated statistic.

[0145] [Advantages of the Invention Embodiments] ●The embodiment uses Dijkstra algorithm or A * We present a method to estimate the weights of road segments for optimal path finders in graphs, such as algorithms. In the context of EV, optimal means more energy efficient. ● Embodiments may transform an inaccessible system of simultaneous equations into a tractable system. This is possible when the components of the aggregated statistics form clusters. ● The embodiments may be applied to any problem where it is necessary to find the contribution of each individual component to an aggregated statistic. For example, the embodiments may enable robust energy consumption estimation for each road segment from aggregated statistics over a trip, without requiring link-level ground truth values ​​for energy consumption or car / driver information. The required aggregated statistics can be measured by inexpensive sensors, saving time and money for businesses. Aggregated data is faster to process, thus saving processing time. o Saves money by eliminating the need to obtain expensive detailed data at intermediate points in the aggregate volume being measured. ● The embodiments are scalable and thus may be used for on-the-fly accurate energy calculations and, optionally, for route optimization as well: only a lookup from the dictionary is required to get the appropriate cluster value for a particular component. ● The embodiment has domain transferability, i.e. training a model in one domain is sufficient to make inferences in some other domain. This is due to the clustering of the predictor space. To obtain the energy consumption of other domains, the trips are just clustered accordingly (the total number of clusters should be the same), and then the previously trained model is queried for each cluster. There is no need to train a model from scratch. ● The weights of the fitted model may have a physical meaning after constrained optimization. They represent the energy consumption rate of a road segment. This makes it easier to understand and interpret the predictions made by the neural network. Such understanding is useful, for example, when the neural network has to make important decisions (e.g. in medicine or autonomous driving).

[0146] [Application field] One application may be as part of energy calculations for planned trips, such as Distance to Empty (DTE) for electric vehicles or energy efficient route optimization software that can predict the most energy efficient route as well as energy consumption and remaining driving distance. The embodiments may be used by commercial vehicles to save money by applying energy efficient route estimations, and even by delivery vehicles to minimize costs while completing deliveries on time. They may also be used by ride sharing companies to optimize pickup locations and maximize profits.

[0147] The embodiments may also be used in scenarios where battery capacity is limited (e.g., used EVs with small battery capacity). Drivers tend to purchase EVs with large battery capacity due to uncertainties in energy consumption during a trip (e.g., due to weather, traffic congestion, etc.). If such uncertainties were reduced, more drivers may purchase EVs with small battery capacity because they are cheaper.

[0148] Another area where embodiments may be applied is vehicle-to-grid (V2G) in vehicular scenarios. With the increase in natural disasters (e.g., high winds, floods, earthquakes), affected areas sometimes suffer secondary effects from drought. EVs can be used as mobile generators in such situations. As EVs use electricity to travel to affected areas, it is important to select the most energy-efficient route to effectively utilize EVs as mobile generators.

[0149] Embodiments may be directed to time series data rather than spatial data. Time series suitable embodiments may be applied, for example, to photovoltaic (PV) energy generation estimation or financial forecasting using aggregated statistics.

[0150] More generally, the modeling approach may be used for any problem that requires estimation of the individual components that contribute to an aggregated statistic.

[0151] [Hardware Implementation] FIG. 14 is a schematic diagram of how data is used in the deployment. A database including geographical data and manufacturer data provides both the cluster modeling (which may be performed, for example, on the cloud) and the route search engine (which may use a laptop computer or a mobile device to view the data formed by the modeling) that suggests the lower / lowest energy route. In one example, a user of the route search engine inputs a start and end point, and the route search engine device queries a dictionary / lookup table provided by the cluster model to segment the various possible routes between those points and determine the rate of energy use of each segment or the total energy use. The route search engine may then compare the various routes to find the one that is the most energy efficient. Alternatively, the segmentation may be performed remotely from the route search engine. As will be clear to those skilled in the art, there are various ways to perform the method on the cloud, off the cloud, or partly locally and partly on the cloud.

[0152] For example, to generate the models (and therefore the clusters), the method requires geographical and manufacturer data. This modeling may be performed offline and the cluster weights may be stored in a look-up table, as represented in Fig. 14 by the arrows connecting the "Geographical Data" and "Manufacturer Database" units to the "Cluster Modeling" unit. Furthermore, the "Geographical Data" unit may be used on the fly if there is real-time traffic data available.

[0153] As a pre-processing step, the method may query the shortest route between the start and end points of each trip and keep the trips that deviate less than 2% from the shortest route. To do so, the method may use manufacturer data about the trip (latitude and longitude of the start and end of the trip) and a route search engine, as represented in Figure 14 by the arrow connecting the "Manufacturer Database" to the "Route Search Engine" unit.

[0154] Finally, the "cluster modeling" unit and the "route search engine" unit may communicate in two different ways. First, an offline dictionary with the mapping of {road segments<->clusters} and {clusters<->weights} is stored in memory. Then, after the user enters a query on the fly (start and end points of the trip), the "route search engine" unit uses the clusters formed by the method to obtain the weights of each road segment in the map, and then uses a shortest graph algorithm (e.g. Dijkstra's algorithm or A. * ) may be used. Thus, the "Route Search Engine" unit may first query the database for clusters and weights (arrow from "Route Search Engine" unit to "Cluster Modeling" unit) and then receive the information (arrow from "Cluster Modeling" unit to "Route Search Engine") to perform energy-efficient route optimization.

[0155] Figure 15 is a block diagram of a computing device, such as a data storage server, that embodies the invention and may be used to implement an embodiment method for predicting energy usage for a route, or a method for modeling a cluster, and / or a method for finding an energy optimized route, as shown in Figure 2, as defined in the claims and described herein. It may also be used in other implementations that estimate the individual components that contribute to an aggregated statistic, as described herein. The computing device has a processor 993 and a memory 994. Optionally, the computing device also includes a network interface 997 for communication with other computing devices, e.g., other computing devices of the invention embodiments.

[0156] For example, an embodiment may consist of a network of such computing devices. Optionally, the computing devices further include one or more input mechanisms 996, such as a keyboard and mouse, and one or more display units 995, such as a monitor. The components can be connected to each other via a bus 992.

[0157] The memory 994 may include a computer-readable medium. This may refer to a single medium or multiple media (e.g., centralized or distributed databases and / or associated caches and servers) on which data structures are stored or configured to carry computer-executable instructions. Computer-executable instructions may include, for example, instructions that can be accessed by a general purpose computer, a special purpose computer, or a special purpose processing device (e.g., one or more processors) to cause them to perform one or more functions or operations. Thus, the term "computer-readable storage medium" may include any medium capable of storing, encoding, or carrying a set of instructions for execution by a machine, causing the machine to perform any one or more of the methods disclosed herein. The term "computer-readable storage medium" may therefore be considered to include, but is not limited to, solid-state memory, optical media, and magnetic media. By way of example, and not limitation, such computer-readable media may include non-transitory computer-readable storage media including random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), compact disk read-only memory (CD-ROM) or other optical disk storage, magnetic disk storage or other magnetic storage devices, flash memory devices (e.g., solid-state memory devices).

[0158] The processor 993 is configured to control the computing device and perform processing operations, such as executing code stored in memory to implement various different functions described herein. For example, the processor 993 may perform steps of matching road segments to corresponding clusters and thus predicted energy usage. Additionally or alternatively, the processor 993 may perform steps of predicting energy usage for a new route by adding up, for each segment on the route, the predicted energy usage multiplied by the length of the segment.

[0159] The memory 994 stores data that is read and written by the processor 993. For example, it may include a database as referred to herein, or it may simply store parameters such as predictor levels and tables of information, as shown in Figures 9, 10, 11, and 12. As referred to herein, the processor 993 may include one or more general-purpose processing devices, such as a microprocessor, a central processing unit, or the like. The processor 993 may include a complex instruction set computing (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, or a processor implementing other instruction sets or a processor implementing a combination of instruction sets. The processor 993 may also include one or more special-purpose processing devices, such as an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a digital signal processor (DSP), a network processor, or the like. In one or more embodiments, the processor 993 is configured to execute instructions to perform the operations and steps described herein.

[0160] The display unit 995 may display a representation of data stored by the computing device, and may also display cursors and dialog boxes and screens that allow interaction between a user and programs and data stored on the computing device. An input mechanism 996 may allow a user to input data and instructions into the computing device. In one example, the display may be used to show the model that was built and the inputs used for the hyperparameters. Additionally or alternatively, the display may show a map including the proposed route and potentially other, less energy efficient routes, and the inputs may be the start and end of the route. In a more general embodiment, the display may show one or more predictions of contributions to an aggregated statistic incorporating one or more segments from the K trials.

[0161] The network interface (network I / F) 997 may be connected to a network such as the Internet and to other such computer devices via the network to enable the computer device to access a manufacturer's database or, in a more general embodiment, a K-trial database, for example. The network I / F 997 may control data input from or output to other devices via the network. Other peripheral devices such as microphones, speakers, printers, power supply units, fans, cases, etc. may be included in the computer device.

[0162] A method embodying the present invention may be performed on a computing device such as that depicted in Figure 15. Such a computing device need not have all of the components depicted in Figure 15, but may consist of a subset of those components. A method embodying the present invention may be performed by a single computing device in communication with one or more data storage servers over a network. The computing device may be a data storage that itself stores the predicted energy usage of a route.

[0163] Methods embodying the invention may be performed by a number of computing devices cooperating with one another, one or more of which may be data storage servers that store at least a portion of the data, such as the predicted energy usage of a cluster or the predicted energy usage of a route.

[0164] In addition to the above embodiments, the following supplementary notes are disclosed. (Appendix 1) 1. A computer-implemented method for forecasting energy usage for a route, comprising: inputting map data for roads included in K trips within a geographic region and inputting a predictor of rate of energy usage along said roads; inputting energy consumption data for the K trips, the energy consumption data indicating a total energy usage T between a start point A and an end point B of each of the K trips; Let each of the roads in the map data for all of the K trips be a length index λ i and Grouping the segments from the trip into a number N of clusters, the clusters being defined according to the range of at least one of the predictors, each cluster having a weight W to be determined. j and The weights W are calculated based on solving one simultaneous equation for each trip. jusing an algorithm to build a model predicting the total energy usage T of the trip, each equation equating the known total energy usage T of the trip to the sum of the known length indices of each segment in the trip multiplied by the weight of the cluster into which the segments are grouped; For each segment, the predicted weight W applied to the cluster into which the segment is grouped j the predicted rate of energy usage Y i and a rate Y of the predicted energy usage to enable a prediction of energy usage for a route within the geographic region incorporating one or more of the segments. i The segment ID is stored along with the instruction The method according to claim 1, (Appendix 2) Before inputting the map data, for each trip in the database, find the shortest route between the start point A and the end point B, and check the length in the database between the start point A and the end point B, and include the shortest route data of the trip if its length is within a threshold of the shortest route between the start point A and the end point B to form the K trips. The method described in Appendix 1. (Appendix 3) and further comprising data cleaning prior to inputting the map data. 3. The method according to claim 1 or 2. (Appendix 4) and cleaning the data includes flattening the elevation data on an elevation tile by modifying an elevation of one or more edge points on the tile that are adjacent to a tile of a different elevation, if there is more than one point on the tile, to bring the elevation of the one or more edge points closer to the elevation of the tile of the different elevation. The method described in Appendix 3. (Appendix 5) The method comprises: dividing the portions of the road within the geographic area that are not included in the trip into unexplored segments; grouping any unexplored segments into clusters according to a level of at least one of said predictors in the unexplored segments; using a length indicator of each unexplored segment and said predicted rate of energy usage to enable a prediction of energy usage for a route incorporating one or more of said unexplored segments; to include all said segments within said geographic region by 5. The method according to any one of claims 1 to 4. (Appendix 6) The predictors include at least one of a gradient in the segment, such as an average positive or negative gradient in the segment, and features that affect traffic in the segment, such as historical traffic data, traffic signals, junctions, and stores. 6. The method according to any one of claims 1 to 5. (Appendix 7) the clusters are defined according to ranges of one or more predictors resulting in a predictor space with a number of dimensions equal to the number of predictors; The predictor space may, for example, have one axis for each predictor, Predictors that have a large impact on the energy usage are preferably divided into more cluster divisions than predictors that have a small impact on the energy usage. 7. The method according to any one of claims 1 to 6. (Appendix 8) For each dimension, a partition representing a level of a predictor having a smaller impact on energy consumption is provided at a lower partition and / or at a lower position along said axis for that predictor than a partition representing a level of a predictor having a larger impact on energy consumption; j W j+1 and ordering the partitions of the clusters so that the partitions of the clusters have a lower value than The method described in Appendix 7. (Appendix 9) The number M of road segments in the geographical area is at least one order of magnitude greater than the number N of clusters and / or the number K of trips; 9. The method according to any one of claims 1 to 8. (Appendix 10) the algorithm for constructing the model is a quadratic optimization algorithm or is developed by a neural network; 10. The method according to any one of claims 1 to 9. (Appendix 11) the algorithm is a quadratic optimization algorithm constrained with constraints on the weights, the weights being constrained according to the clusters and ordered according to the influence each predictor has on the energy consumption rate; 11. The method according to claim 10. (Appendix 12) 1. A computer-implemented method for route optimization, comprising: using a predicted energy usage of possible routes in said geographical area between a start point and an end point, calculated according to the method of any one of claims 1 to 11; Indicating to the user the relative energy usage of the possible routes, for example by showing the lowest energy route; The method according to claim 1, (Appendix 13) A computer program comprising instructions which, when executed by a computer, cause the computer to carry out the method according to claim 1. (Appendix 14) 13. A data processing apparatus having a processor and memory configured to carry out the method of any one of claims 1 to 12. (Appendix 15) 1. A computer-implemented method for estimating individual components contributing to an aggregated statistic, comprising: inputting data for K trials and inputting a predictor of overall percentage contribution to the aggregated statistic; inputting contribution data for the K trials, the contribution data indicating a total contribution T between a start and end of each of the K trials; Let each of the K trials be a length index λi and dividing the time or space into segments of Grouping the segments from the K trials into a number N of clusters, the clusters being defined according to the range of at least one of the predictors, each cluster having a weight W j and The weights W are calculated based on solving one simultaneous equation for each trial. j using an algorithm to build a model predicting T, where each equation equates the known total contribution T of the trial to the sum of the known length indices of each segment in the trial multiplied by the weight of the cluster into which the segments are grouped; For each segment, the predicted weight W applied to the cluster into which the segment is grouped j Let Y be the predicted contribution ratio. i and said predicted contribution proportion Y i The segment ID is stored along with the instruction The method according to claim 1, [Explanation of symbols]

[0165] 992 Bus 993 Processor 994 Memory 995 Display Unit 996 Input Mechanism 997 Network Interface

Claims

1. 1. A computer-implemented method for forecasting energy usage for a route, comprising: inputting map data of roads included in K trips within a geographic region into a machine learning model, and inputting variables into the machine learning model for predicting a ratio of energy use along the roads to energy use of the route; inputting energy consumption data of the K trips into the machine learning model, the energy consumption data indicating a total energy usage T between a start point A and an end point B of each of the K trips; Each of the roads in the map data for all of the K trips is represented by an index λ ｉ and dividing the signal into segments represented by grouping the segments from the trip into a number N of clusters, the clusters being defined according to the range of at least one of the variables, each cluster having a weight W to be determined according to the impact of the at least one variable on energy usage; ｊ and The weights W are calculated based on solving one simultaneous equation for each trip. ｊ using an algorithm to build a model predicting the total energy usage T of the trip, each equation equating the known total energy usage T of the trip to the sum of known length indicators of each segment in the trip multiplied by the weight of the cluster into which the segments are grouped; For each segment, the predicted weight W applied to the cluster into which the segment is grouped ｊ Let Y be the predicted rate of energy usage. ｉ and a rate of said predicted energy usage Y for use in predicting energy usage for a route within said geographic region incorporating one or more of said segments. ｉ storing a segment ID for identifying the segment together with the instruction; The method according to claim 1,

2. forming the K trips by, before inputting the map data, for each trip in a database, finding a shortest route between the start point A and the end point B, checking the length of the trip in the database between the start point A and the end point B, and determining to consider the trip if a difference in length between the shortest route and the trip is within a predetermined range; The method of claim 1.

3. and further comprising data cleaning prior to inputting the map data. The method according to claim 1 or 2.

4. The map data includes elevation mesh tiles, and the data cleaning comprises: for each tile of the elevation mesh tiles, if there is more than one point on the tile, flattening the elevation data on the elevation mesh tiles by modifying an elevation of one or more edge points on the tile that are adjacent to a tile of a different elevation to bring the elevation of the one or more edge points closer to the elevation of the tile of the different elevation. The method according to claim 3.

5. The method comprises: dividing the portions of the road within the geographic area that are not included in the trip into unexplored segments; grouping any unexplored segments into clusters according to the range of at least one of said variables in the unexplored segments; using an indication of the length of each unexplored segment and said predicted rate of energy usage to generate a prediction of energy usage for a route incorporating one or more of said unexplored segments; to include all said segments within said geographic region by 5. The method according to any one of claims 1 to 4.

6. The variables include at least one of a gradient in the segment, such as an average positive or negative gradient in the segment, and features that affect traffic in the segment, including historical traffic data, traffic signals, junctions, and stores.

6. The method according to any one of claims 1 to 5.

7. the clusters are defined according to ranges of one or more variables resulting in a space with a number of dimensions equal to the number of the variables; The space is a coordinate space having coordinate axes corresponding to the dimensions of each variable, The variables that have a large impact on the energy use are divided into more cluster divisions than the variables that have a small impact on the energy use, and the impact of each variable on the energy use is determined automatically using a random forest algorithm.

7. The method according to any one of claims 1 to 6.

8. ordering the partitions of the cluster such that, for each dimension, partitions representing levels of variables with less impact on energy consumption are associated with lower weights than weights associated with partitions representing levels of variables with more impact on energy consumption; The method of claim 7.

9. the number M of road segments in the geographical area is at least an order of magnitude greater than the number N of clusters and / or the number K of trips; 9. The method according to any one of claims 1 to 8.

10. the algorithm for constructing the model is a quadratic optimization algorithm or is developed by a neural network; 10. The method according to any one of claims 1 to 9.

11. the algorithm is a quadratic optimization algorithm constrained with constraints on the weights, the weights being constrained according to the clusters and ordered according to the influence each variable has on the energy consumption rate; The method of claim 10.

12. 1. A computer-implemented method for route optimization, comprising: - using a predicted energy usage of possible routes in said geographical area between a start point and an end point, calculated according to the method of any one of claims 1 to 11; showing the user the relative energy usage of the possible routes by showing the lowest energy route; The method according to claim 1,

13. A computer program comprising instructions which, when executed by a computer, cause the computer to carry out the method of claim 1.

14. 13. A data processing apparatus having a processor and a memory configured to carry out a method according to any one of claims 1 to 12.

15. 1. A computer-implemented method for estimating individual components contributing to an aggregated statistic, comprising: inputting data for K trials into a machine learning model and inputting variables into the machine learning model for predicting the percentage contribution of the individual components to the aggregated statistic; inputting contribution data of the K trials into the machine learning model, the contribution data indicating a total contribution T between a start and end of each of the K trials; Each of the K trials is expressed as a length of an index λ ｉ and dividing the time or space into segments represented by Grouping the segments from the K trials into a number N of clusters, the clusters being defined according to the range of at least one of the variables, each cluster having a weight W, which should be determined according to the influence of the at least one variable on the statistic. ｊ and The weights W are calculated based on solving one simultaneous equation for each trial. ｊ using an algorithm to build a model predicting T, where each equation equates the known total contribution T of the trial to the sum of known length indices of each segment in the trial multiplied by the weight of the cluster into which the segments are grouped; For each segment, the predicted weight W applied to the cluster into which the segment is grouped ｊ Let Y be the predicted contribution ratio. ｉ and for use in predicting the contribution for an aggregate statistic incorporating one or more of the segments, ｉ storing a segment ID for identifying the segment together with the instruction; The method according to claim 1,

Images