Method for predicting power consumption in electrical network

Abstract

The embodiment of the invention relates to a method for predicting power consumption in a power grid. The method employs a linear autoregression model to calculate prediction data related to the power consumption in the power grid. The prediction method ensures a high level of performance in terms of prediction accuracy and can be easily implemented even if limited computing and data storage resources are available.

Description

Technical Field

[0001] The present invention relates to the field of power distribution networks. More specifically, the present invention relates to a method for predicting power consumption in a power grid. Background Art

[0002] As is known to all, the management of the electric grid generally requires accurate forecasting of electric energy consumption to allow system operators to correctly plan the use of electric energy over time, thereby preventing or limiting demand peaks and formulating more favorable electric energy purchasing plans.

[0003] The most common forecasting methods are based on machine learning (ML) techniques and require processing of relevant amounts of data to provide accurate forecasts. In addition, these methods typically provide for the execution of a computationally intensive training phase of an artificial intelligence unit (e.g., a neural network) that executes the ML algorithm.

[0004] Therefore, ML-based prediction methods are difficult to implement through computing systems that usually manage the operation of field devices and switchboards in the power grid, which usually have relatively limited storage and computing resources and are therefore not suitable for processing large amounts of data. In fact, these computing systems are usually based on edge computing architectures, and their basic purpose is to bring computing and data storage closer to the source of data to shorten response times and save bandwidth, rather than processing large data sets.

[0005] In the prior art, prediction methods have been developed (e.g., based on linear regression analysis techniques), which generally require lighter computing and data storage resources than ML-based prediction methods and are therefore suitable for implementation in computing systems typically used to manage power grids.

[0006] An example of these prediction methods is described in US10515308B2.

[0007] However, this type of available forecasting methods tend to perform relatively poorly in terms of reliability and forecasting accuracy compared to ML-based forecasting techniques. Summary of the invention

[0008] The main task of the present invention is to provide a method for predicting electric energy consumption in a power grid, which can overcome the limitations of the above-mentioned prior art.

[0009] Within this aim, another object of the present invention is to provide a prediction method which can ensure a high level of performance in terms of reliability and prediction accuracy.

[0010] Another object of the present invention is to provide a prediction method that can be easily implemented even when limited computing and data storage resources are available, and is therefore suitable for implementation in computing systems typically used to manage power grid operations, such as in computing systems based on edge computing architectures.

[0011] According to the invention, this task and these objects, as well as other objects which are clear from the subsequent description and the accompanying drawings, are achieved by a prediction method according to claim 1 and the related dependent claims set out below.

[0012] In a general definition, the method according to the invention comprises the following steps:

[0013] - Obtaining first detection data, the first detection data including the actual

[0014] Detection values ​​related to power consumption;

[0015] - acquiring additional detection data comprising detection values ​​related to energy consumption in the above mentioned power grid during at least a time window before a given reference moment;

[0016] - obtaining calendar data including time periods associated with operation of the electrical network

[0017] time information;

[0018] - calculating training data based on the acquired detection data and the acquired calendar data;

[0019] - setting a linear autoregressive mathematical model describing the trend of the electric energy consumption in the electric network based on the training data. Such a linear autoregressive mathematical model is configured to process at least a set of exogenous input values ​​indicating at least a periodic function approximating the trend of the electric energy consumption in the electric network in the at least time window before the reference time.

[0020] Energy consumption curve;

[0021] - calculating forecast data based on the above-mentioned linear autoregressive model, the forecast data comprising forecast values ​​related to the consumption of electrical energy in the above-mentioned power grid during a time window after the above-mentioned reference moment.

[0022] Preferably, the method according to the invention comprises the step of acquiring second detection data comprising detection values ​​related to energy consumption in the above-mentioned power grid during a first time window before the above-mentioned reference moment. In this case, the linear autoregressive mathematical model is configured to process first exogenous input values ​​indicative of a first periodic function that approximates a curve of electrical energy consumption in the above-mentioned power grid during the above-mentioned first time window.

[0023] Preferably, the method according to the present invention further comprises the step of acquiring third detection data, the third detection data comprising detection values ​​related to energy consumption in the above-mentioned power grid during a second time window before the above-mentioned reference moment. In this case, the linear autoregressive mathematical model is configured to process a second exogenous input value indicating a second periodic function, which approximates a curve of electrical energy consumption in the above-mentioned power grid within the above-mentioned second time window. Preferably, such a second time window is longer than the above-mentioned first time window. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Other features and advantages of the present invention will appear more clearly from the description of a preferred but non-exclusive embodiment thereof, shown purely by way of example and not by way of limitation in the accompanying drawings, in which

[0025] Figure 1 schematically illustrates a common power grid for power distribution applications, and

[0026] Figure 2-Figure 5 The prediction method according to the present invention is schematically illustrated. DETAILED DESCRIPTION

[0027] With reference to the above-mentioned drawings, the present invention relates to a method for predicting a power grid 1 ( Figure 1 ) in a method 100 for power consumption.

[0028] In principle, the power grid 1 may be of any type, such as a smart grid, a microgrid, or a power distribution network for industrial, commercial or residential buildings or factories.

[0029] Preferably, the power grid 1 operates at a low voltage or medium voltage level, wherein the term "low voltage" relates to operating voltages up to 1.2 kV AC and 1.5 kV DC, and the term "medium voltage" relates to operating voltages above 1.2 kV AC, 1.5 kV DC, up to several tens of kV (e.g. up to 72 kV AC and 100 kV DC).

[0030] The power grid 1 may be single-phase or multi-phase (eg, three-phase). Generally, the power grid 1 may be electrically connected to one or more power sources 2 (eg, power companies) and one or more electrical loads 3, each of which consumes a corresponding amount of electrical energy in operation.

[0031] The power grid 1 may include one or more field devices 4 (e.g., switching devices, sensors, etc.) configured to regulate the flow of power along a branch of the power grid, and one or more intelligent electronic devices 5 (e.g., controllers, protective relays, smart interfaces, etc.) configured to control the operation of the above-mentioned field devices and more generally control the operation of the power grid.

[0032] Advantageously, the intelligent electronic device 5 may be equipped with suitable computing and storage resources to process data related to the operation of the electrical grid.

[0033] Preferably, the intelligent electronic device 5 is based on an edge computing architecture.

[0034] Generally, the above-mentioned power grid 1, power source 2, electrical load 3, field device 4 and intelligent electronic device 5 may be of known types, and for the sake of brevity, will not be described in further detail here.

[0035] The method 100 according to the invention is suitable for being performed by a computerized device. The computerized device may advantageously comprise data processing resources capable of executing software instructions configured to implement the method.

[0036] Such a computerized device is preferably an intelligent electronic device 5 of the power grid, which can be installed in the field as a stand-alone device (e.g., a controller) or embedded in an electrical device 4 (e.g., as a protection relay). For example, such an intelligent electronic device can be an intelligent switchboard HMI, which is operably coupled to a certain number of field devices 4 of the power grid and is configured to process data related to the operation of the power grid.

[0037] As will be better explained below, the method 100 is implemented by using a special mathematical model M R To calculate the forecast data D related to the power consumption in the power grid P , the mathematical model M R Based on the training data D T To loop setting, the training data D T By collecting and processing real-time and historical energy consumption data in the power grid S1 , D S2 , D S3 to calculate.

[0038] refer to Figure 2 , in setting the subsequent reference time t of the above mathematical model R Continuously obtain real-time detection data D S1 , and obtain historical detection data D S2 , D S3 . Historical test data D S2 , D S3 It refers to the time reference time t R The different training time windows TW1 , TW2 are previously described and have different selectable durations (eg weeks, months or years, respectively).

[0039] At the corresponding reference time t R After being set, the mathematical model M Ris used to calculate the R The predicted data related to future power consumption in the power grid during a subsequent prediction time window TW3 (eg, one week).

[0040] During the above prediction time window TW3, the above prediction data D P With a predefined time granularity T P (e.g., 15 minutes) and refer to a predefined time horizon T H Calculated periodically (e.g., 24 hours).

[0041] At the end of the prediction time window TW3, a new mathematical model is set at a new reference time, and the mathematical model can be used to calculate the above prediction data within a new prediction time window after the above new reference time.

[0042] The prediction method 100 according to the present invention will now be described in detail.

[0043] refer to Figure 2-Figure 3 The method 100 includes acquiring first detection data D S1 Step 101, the first detection data D S1 Includes detected values ​​related to actual (real-time) electrical energy consumption in the power grid.

[0044] The first detection data D S1 Refers to the instantaneous energy consumption in the power grid at each common operating moment.

[0045] The first detection data D S1 It may be collected from one or more field devices 4 (eg, sensors) or from one or more intelligent electronic devices 5 installed in the field or even from a remote computerized device.

[0046] Preferably, step 101 is performed continuously, possibly in parallel with other steps of method 100. Therefore, the first detection data D S1 A vector comprising detection values, which are continuously and cyclically acquired at subsequent acquisition moments and stored in a memory, wherein two consecutive acquisition moments are separated by a time interval corresponding to a predefined acquisition period of the above-mentioned first detection data.

[0047] According to the present invention, the prediction method 100 comprises one or more acquisition steps 102, 103, wherein one or more sets of additional detection data D are acquired. S2 , D S3 .

[0048] Each set of additional data includes the data at a given reference time t R The detection values ​​related to the historical energy consumption in the power grid during the previous corresponding time windows TW1 and TW2. As mentioned above, the time reference time tR It is used to calculate the forecast data related to the future power consumption in the power grid. P Mathematical model M R moment.

[0049] Preferably, the prediction method 100 includes acquiring the second detection data D S2 Step 102, the second detection data D S2 Including at the reference time t R Detection values ​​related to historical power consumption in the power grid during the previous first time window TW1.

[0050] The second detection data D S2 It refers to the time reference moment t R Energy consumption in the grid at the operating instant preceding and included in the first time window TW1.

[0051] Each time the mathematical model is established, the duration of the first time window TW1 may be selected. For example, the first time window TW1 may refer to the time reference time t R The weeks before.

[0052] The second detection data D S2 It may be retrieved from a memory or from one or more intelligent electronic devices 5 installed on site or even from a remote computerized device.

[0053] Preferably, the prediction method 100 includes acquiring the third detection data D S3 Step 103, the third detection data D S3 Including at the reference time t R Detection values ​​related to historical power consumption in the power grid during the previous second time window TW2.

[0054] The third detection data D S3 It refers to the time reference moment t R The energy consumption in the grid at the operating instant preceding and included in the second time window TW2.

[0055] The duration of the second time window TW2 is preferably much longer than the duration of the first time window TW1. For example, the second time window TW2 may refer to the time reference instant t R Months or years before.

[0056] The third detection data D S3 It may be retrieved from a memory or from one or more intelligent electronic devices 5 installed on site or even from a remote computerized device.

[0057] According to the present invention, the prediction method 100 comprises obtaining calendar data D CStep 104, the calendar data D C Contains time information relevant to the operation of the power grid. Calendar data D C Information relating to working days, non-working days or holidays, or more generally information relating to other time conditions that may affect energy consumption of the power grid may be included.

[0058] The collected time information advantageously refers to the reference time t R The following time window is TW3.

[0059] Calendar data D C It may be retrieved from a memory or from one or more intelligent electronic devices 5 installed on site or even from a remote computerized device.

[0060] According to the present invention, the prediction method 100 comprises calculating the training data D by processing the acquired detection data and calendar data. T Step 105.

[0061] Training data D T It is intended to be used to set up a mathematical model that describes the R The trend of the electrical energy consumption in the grid during the subsequent third time window TW3.

[0062] In principle, the duration of the third time window TW3 may vary each time the above mathematical model is established. For example, the third time window TW3 may refer to the time reference time t R The week after.

[0063] Preferably, the calculation step 105 includes processing the first detection data D continuously acquired from the external data source S1 , to check the correctness of the acquired data. Advantageously, the first detection data D S1 The data are processed with the help of suitable statistical techniques to identify outliers or missing values. Possible incorrect detected values ​​can be conveniently replaced by using suitable interpolation techniques.

[0064] Preferably, the calculation step 105 includes processing the acquired second detection data D S2 , to identify the trend of power consumption in the power grid during the first time window TW1. In practice, the second detection data D acquired by appropriate statistical techniques is analyzed. S2 , to indicate the reference time t R The short-term behavior of electricity consumption in the grid before.

[0065] As will become clearer from the following, the information obtained by the processing activity is conveniently used to calculate a first periodic function U(t) describing the first periodic function U(t) in the first time window WT1 Distribution of electrical energy consumption in the intra-grid.

[0066] This allows to set up the above mathematical model M in a way that takes into account possible short-term nonlinearities affecting the energy consumption in the grid. R .

[0067] The information obtained by this processing activity can also be used to appropriately tune the duration of the first time window TW1.

[0068] For example, if the detected power consumption shows a high level of periodicity or a less regular curve, respectively, the first time window TW1 may be set at the reference time t R Tune in two or four weeks before.

[0069] Advantageously, the duration of the first time window TW1 can also be adjusted based on time information with reference to the above-mentioned time window, which time information can advantageously be obtained from the acquired calendar data D C It can be concluded.

[0070] Preferably, the calculation step 105 includes processing the acquired third detection data D S3 , to identify the trend of power consumption in the power grid during the second time window TW2. In practice, the third detection data D is obtained by analyzing the acquired third detection data D through appropriate statistical techniques. S2 , to indicate the reference time t R Previous long-term trend (or seasonality) of the electrical energy consumption in the grid. This is not possible if only historical detection data related to the first time window TW1 is considered.

[0071] As will become clearer from the following, the information obtained through this processing activity is conveniently used to calculate the second periodic function U l (t), the second periodic function U l (t) Curve describing the energy consumption in the grid during the second time window TW2.

[0072] This allows setting up the above mathematical model in a way that takes into account possible long-term non-linear or seasonal factors affecting the electrical energy consumption in the grid.

[0073] According to the present invention, the prediction method 100 comprises: T Set up the linear autoregression mathematical model M R Step 106.

[0074] Mathematical Model M R describes the trend of electrical energy consumption in the grid and is intended to be used to calculate the energy consumption at the reference time t R Prediction data D related to power consumption at the next time P, in particular at a subsequent time k included in the power grid during the third time window TW3.

[0075] Due to its autoregressive nature, at each time k+1, the mathematical model M R is configured to process endogenous input values ​​y(k) associated with the previous time instant k.

[0076] The endogenous input value y(k) includes previously calculated predicted values ​​related to the power consumption in the power grid, and may also include the first detection data D S1 Previously acquired detection values ​​related to the instantaneous power consumption in the power grid included.

[0077] However, according to a particularly important aspect of the present invention, the mathematical model M R is configured to process one or more sets of exogenous input values ​​U(k), U l (k).

[0078] Each set of exogenous input values ​​indicates the corresponding periodic function U(t), U l (t), the periodic functions U(t), U l (t) is approximately equal to the reference time t R The curves of the power consumption in the power grid in the previous corresponding time windows TW1 and TW2.

[0079] Preferably, the linear autoregressive model M R is configured to process the reference time t R The first exogenous input value U(k) associated with the previous time instant k.

[0080] The first exogenous input value U(k) indicates a first periodic function U(t) which approximates the profile of the electrical energy consumption in the grid within the first time window TW1 .

[0081] Preferably, the first periodic function U(t) is a combination of cosine and sine functions with unit amplitude and different frequencies, for example ranging from hourly values ​​to weekly values. Advantageously, the first periodic function U(t) is based on the second detection data D acquired by processing S2 The acquired training data is used to calculate to identify the trend of the power consumption in the power grid during the first time window TW1.

[0082] Therefore, at the reference time t R The vector of the first exogenous input value U(k) at the previous universal time instant k can be expressed as a combination of n sinusoidal terms according to the following expression:

[0083] U(k)=[cos(w1k),...,cos(w n k), sin(w1k), ..., sin(wn k)]

[0084] Among them, the terms w1, ..., w n Indicates the frequency selected in order to approximate the curve of the electrical energy consumption within the first time window TW1.

[0085] Preferably, the linear autoregressive model M R is configured to process the R The second exogenous input value U related to the previous time k l (k).

[0086] The second exogenous input value U l (k) indicates the second periodic function U l (t), the second periodic function U l (t) is a curve approximating the energy consumption in the grid during the second time window TW2.

[0087] Preferably, the second periodic function U l (t) is a combination of cosine and sine functions with unit amplitude and different frequencies, for example ranging from monthly values ​​to annual values.

[0088] Advantageously, the second periodic function U l (t) Based on the third detection data D obtained by processing S3 The acquired training data is used to calculate to identify the trend of power consumption in the power grid during the second time window TW2.

[0089] Therefore, at the reference time t R The second exogenous input value U at the previous universal time k l The vector of (k) can be expressed as a combination of q sinusoidal terms according to the following expression:

[0090] U l (k) = [cos(w1k), ..., cos(w q k), sin(w1k), ..., sin(w q k)]

[0091] Therein the terms w1, . . . , wq indicate frequencies selected in order to approximate the distribution of the electrical energy consumption within the second time window TW2.

[0092] Mathematical Model M R The setting can be conveniently performed in a training phase, which can include one or more training steps.

[0093] At each training event, the parameters θ of the mathematical model are iteratively calculated until a maximum number of training steps is reached or the estimation error of the calculated prediction value is sufficiently low.

[0094] Set the mathematical model M R Step 106 conveniently comprises setting the autoregressive order m of the mathematical model and setting the maximum number T of training steps for training the mathematical model. max .

[0095] The autoregressive order m can be chosen based on the desired model complexity level or the time window TW considered. i The length of D can be selected based on the parameters selected in the cross-validation phase performed before implementing the above method, for example, in different datasets D S0 on (if available).

[0096] For example, the autoregressive order m may be set to m=3.

[0097] The maximum number of training steps T max is a parameter that can be chosen depending on, for example, the performance of available edge computing units or possible time constraints of the considered application.

[0098] Preferably, the number of training events is T>1. In this case, two subsequent training events are advantageously separated by a time interval corresponding to the time interval used to calculate the prediction data D P The time interval is relatively long compared to the time granularity set by the training event. For example, if a time granularity of 15 minutes is set, the time interval between two subsequent training events can be 24 hours.

[0099] Preferably, the linear autoregressive mathematical model M R is a linear ARX mathematical model with one or more (more preferably multiple) exogenous inputs.

[0100] Generally speaking, the mathematical model M R It can be expressed as:

[0101]

[0102] where k = m + 1, ..., T H / T P .

[0103] in:

[0104] -y(k) is a vector of endogenous input values. For k <= 2m, y(k) includes the previously calculated prediction values ​​and the first detection data D S1 The real-time detection value included in y(k) only includes

[0105] The previously calculated predicted values ​​for k>2m;

[0106] U(k) is a vector of first exogenous input values ​​indicating a first periodic function U(t) approximating the energy consumption in the grid during the first time window TW1.

[0107] Consumption curve;

[0108] -U l (k) is the second periodic function U l (t) is a vector of the second exogenous input values, the second periodic function U l (t) is approximately equal to the power consumption in the power grid in the second time window TW2.

[0109] Consumption curve;

[0110] - θ is the model parameter vector to be calculated during the training phase of the mathematical model;

[0111] -T H is set to calculate the predicted data D P For example, T H Can

[0112] Set to T H =24 hours;

[0113] -T P is set to calculate the predicted data D P For example, T P Can be set to T P =15 minutes.

[0114] -Ratio T H / T P defines the number of predicted and detected values ​​to be considered for calculating the prediction error. For example, the ratio takes the value T H / T P =96, where the time horizon T H is set to 24 hours, and the time granularity T P It is set to 15 minutes.

[0115] At each training event, a preliminary vector θ' of model parameters is computed by solving an unconstrained linear problem:

[0116]

[0117] in:

[0118] T is a number representing the time (preferably in days) elapsed until the training event, during which the first detection data D S1 has been collected;

[0119] -A iis the number of times (T H / T P -m) vector of detection values, during which the first detection data D S1 have been collected, where m is the set number of regression steps. For example, the vector A i has (96-m) values, where the time horizon T H is set to 24 hours, and the time granularity T P It is set to 15 minutes;

[0120] -Y i is calculated for time unit i (preferably day i) (T H / T P -m) vector of predicted values, during which the data D is detected S1 For example, the vector Y i With (96-m) values, time horizon T H is set to 24 hours, and the time granularity T P It is set to 15 minutes.

[0121] If the maximum number of training steps T is reached max Or the calculated error ||A i -Y i || is lower than the predefined threshold, then the mathematical model M R The training phase is terminated.

[0122] At the end of the training phase, the last calculated vector θ' of model parameters becomes the final vector θ of model parameters of the mathematical model. R is finally set.

[0123] The final vector of model parameters θ can be expressed as:

[0124] θ=[θ y ,θ U ,θ Ul ]

[0125] in:

[0126] -θ y is the vector of model parameters calculated from the linear combination of endogenous input values ​​y(k);

[0127] -θ U is a vector of model parameters calculated by linearly combining the first exogenous input value input U(k);

[0128] -θ Ul is a linear combination of the second exogenous input value input U l (k) is a vector of calculated model parameters.

[0129] According to one aspect of the invention, the mathematical model M is adjusted based on the corresponding parameters of the mathematical model previously calculated. R One or more final parameters of .

[0130] In particular, the second exogenous input value input U is combined l The final model parameters θ of (k) Ul can be based on the corresponding parameters θ' calculated during the training phase Ul and the corresponding parameter θ" calculated previously Ul (ie the corresponding parameters of the previously set mathematical model acquired during a previous training event) to calculate.

[0131] Combined with the second exogenous input value input U l (k) is the vector θ of the final model parameters Ul It can be calculated as:

[0132] θ Ul =(1-α)*θ′ Ul +α*θ″ Ul

[0133] in:

[0134] -θ' Ul is the vector of corresponding parameters calculated during the most recent training event;

[0135] -θ" Ul is the vector of corresponding parameters calculated during the previous training episode;

[0136] -α is a tunable parameter, 0<α<=1.

[0137] The tunable parameter α allows tuning the adaptation speed of the input values ​​(second exogenous input values) according to the acquired detection values ​​indicating the instantaneous electric energy consumption of the electric network, these input values ​​being suitable for taking into account the long-term seasonality of the electric energy consumption.

[0138] Larger values ​​of α result in slower adaptation because more weight is given to the contribution of parameters calculated during previous training episodes, whereas smaller values ​​of α result in faster adaptation because more weight is given to the contribution of parameters set during the most recent training episode.

[0139] According to the present invention, the prediction method 100 comprises a method based on the mathematical model M set in the previous step 106. R To calculate the predicted data D P Step 107.

[0140] Prediction data D P Included are predicted values ​​relating to the electrical energy consumption in the grid during the third time window TW3.

[0141] As shown above, preferably, at the subsequent calculation time k, the predefined time granularity T P (For example, 15 minutes) cyclically calculate the predicted data D P .

[0142] Preferably, at each calculation time k, with a predefined time horizon T H (e.g., 24 hours) Calculate the predicted data D P .

[0143] Preferably, as described above, the prediction method 100 is cyclically repeated at the end of each third time window TW3.

[0144] When the third time window TW3 expires, at the new reference time t R Set up a new linear autoregressive math R Model.

[0145] Therefore, the first detection data D is continuously acquired in each acquisition period. S1 (Step 101 of method 100) at the same time, based on the new reference time t R The above steps 102 - 106 of method 100 are repeated for the calculated new time windows TW1 , TW2 , and TW3 .

[0146] Then, using the new mathematical model M R To calculate the new reference time t R The predicted values ​​related to the power consumption in the power grid during the new time window TW3 thereafter.

[0147] According to one aspect of the present invention ( Figure 3-Figure 4 ), the method 100 includes performing a first checking process to check at a reference time t R The mathematical model M established at R Step 108 of calculating performance.

[0148] The first checking process 108 is to check the R Calculated prediction data D P Whether it corresponds to the detection data D indicating the actual power consumption in the power grid S1 match.

[0149] Preferably, the first checking process 108 comprises comparing the first detection data D S1 and predicted data D P In step 108a, the two data are respectively the last execution time of the first inspection process 108 (or the reference time t if the inspection process is performed for the first time). R) and the time interval (inspection period) between the current execution time of the inspection process.

[0150] Preferably, the first inspection process 108 comprises a step 108b of calculating an error function E, the error function E indicating the collected first detection data D S1 The detection values ​​included in the calculation of the prediction data D P The difference between the predicted values ​​included in .

[0151] The error function E (which may be, for example, a MAPE error function) is used to calculate the predicted data D P The mathematical model M R A measure of guaranteed forecast accuracy.

[0152] The checking process 108 includes the following steps: TH Update the mathematical model M R Step 108c.

[0153] As mentioned above, the mathematical model M R The updating of is performed by re-executing step 106 of method 100. In practice, as described above, the mathematical model M R Updated by forcing a new training event.

[0154] If the error function E does not exceed the threshold error value E TH , then maintain the mathematical model M R , and terminate the first checking process 108.

[0155] Preferably, the first checking process 108 is performed cyclically during the third time window TW3, for example with a 24-hour checking period.

[0156] According to another aspect of the present invention ( Figure 3 and Figure 5 ), the method 100 includes a step 109 of performing a second checking process to check the predicted electrical energy consumption in the power grid.

[0157] The first checking process 108 is to check the R Calculated prediction data D P Whether it falls within the prediction confidence interval.

[0158] Preferably, the second checking process 109 includes processing the calculated prediction data D P With step 109a of calculating a prediction function P indicative of a predicted trend of electrical energy consumption in the grid.

[0159] To calculate the prediction function P, the calculated prediction data D may be processed by suitable statistical techniques of known type.P .

[0160] Preferably, the second checking process 109 comprises: Max or below a predefined minimum confidence value P Min Step 109b of generating an alarm signal AL in the event of a fault.

[0161] Confidence value P max , P min It can be conveniently calculated in the following way: Calculate the first detection data D collected S1 The detection values ​​included in the calculation of the prediction data D P An error function is derived for the difference between the predicted values ​​included in , and the error function is processed by means of suitable statistical techniques of known type.

[0162] If the above prediction function P takes the value of max , P min If the confidence interval is within the defined range, the second checking process 109 is terminated.

[0163] Preferably, the second checking process 109 is performed cyclically during the third time window TW3, for example with a repetitive period of 24 hours.

[0164] The prediction method 100 according to the present invention provides related advantages.

[0165] The prediction method 100 ensures a high level of performance in terms of prediction accuracy.

[0166] From this perspective, the linear autoregressive model M R is configured to process a first exogenous input value U(k) and a second exogenous input value U l (k) The two cases are particularly relevant. In fact, since in calculating the predicted data D P The solution can significantly improve forecast accuracy by properly considering short-term and long-term factors that may affect the trend of electrical energy consumption.

[0167] After the training phase, the model parameters θ are fine-tuned, especially the parameters θ that aim to simulate the long-term seasonality of electricity consumption. Ul Fine-tuning can further improve the performance of the prediction method 100.

[0168] Iterative checking of the accuracy of the calculated prediction data further improves the reliability of the prediction method.

[0169] To confirm the above, experimental tests have shown that the prediction method 100 ensures an accuracy performance that is fully comparable to that provided by known methods of the prior art based on ML algorithms.

[0170] The prediction method 100 is configured to process relatively small data sets. Therefore, it is particularly suitable for implementation in computing systems with limited computing and data storage resources, such as edge computing systems that are commonly used to manage power grid operations.

[0171] Thus, the forecasting method 100 is particularly well suited for implementation using hardware and software resources already installed in the field to manage the operation of an electrical grid.

[0172] Therefore, the prediction method 100 is suitable for implementation in a digital power distribution network (smart grid, microgrid, etc.).

Claims

1. A method (100) for predicting electrical energy consumption in an electrical network (1), the method comprising the following steps: - Obtain (101) first detection data (D S1 ), wherein the first detection data comprises a detection value related to actual power consumption in the power grid; - Obtain (102, 103) additional detection data (D S2 , D S3 ), the additional detection data includes at the reference time (t R ) during at least a time window (TW1, TW2) before and after the detection value related to energy consumption in the power grid; -Get (104) calendar data (D C ), the calendar data comprising time information associated with operation of the power grid; - Calculate (105) training data (D) based on the acquired detection data and calendar data T ); -Based on the training data (D T ), setting (106) a linear autoregressive mathematical model (M) describing the trend of the electric energy consumption in the power grid R ), the linear autoregressive mathematical model (M R ) is configured to process an indication of at least a periodic function (U(t), U l (t)) of at least the set of exogenous input values ​​(U(k), U l (k)), the periodic function is approximately R ) before the time window (TW1, TW2); -Based on the linear autoregressive model (M R ), calculate (107) prediction data (D P ), the prediction data includes at the reference time (t R ) during a time window (TW3) after ).

2. The method according to claim 1, characterized in that The method comprises acquiring (102) second detection data (D S2 ), the second detection data includes at the reference time (t R ) before the first time window (TW1) and the detection value related to the energy consumption in the power grid, wherein the linear autoregressive mathematical model (M R ) is configured to process a first exogenous input value (U(k)) indicative of a first periodic function (U(t)) which approximates the curve of the electrical energy consumption in the power grid within the first time window (TW1).

3. The method according to any one of the preceding claims, characterized in that The method comprises acquiring (103) third detection data (D S3 ), the third detection data includes the data corresponding to the reference time (t R ) before the second time window (TW2), wherein the linear autoregressive mathematical model (M R ) is configured to process the second periodic function (U l (t)) l (k)), the second periodic function approximates the curve of the electrical energy consumption in the power grid within the second time window (TW2).

4. The method according to any one of the preceding claims, characterized in that Calculate the training data (D T The step (105) includes processing the acquired first detection data (D S1 ) to check the correctness of the data.

5. The method according to claim 2, characterized in that: Calculate the training data (D T The step (105) includes processing the acquired second detection data (D S2 ) to identify the trend of the electrical energy consumption in the power grid during the first time window (TW1).

6. The method according to claim 3, characterized in that Calculate the training data (D T The step (105) includes processing the acquired third detection data (D S3 ) to identify the trend of the electrical energy consumption in the power grid during the second time window (TW2).

7. The method according to any one of the preceding claims, characterized in that The predicted data (D P ) with a predefined time granularity (T P ) and a predefined time horizon (T H ) loop calculation.

8. The method according to any one of the preceding claims, characterized in that The linear autoregressive mathematical model (M R ) is a linear ARX mathematical model with one or more exogenous inputs.

9. The method according to any one of the preceding claims, characterized in that Set the linear autoregressive mathematical model (M R )include: - is the linear autoregressive mathematical model (M R ) Set the regression order (m) and the maximum number of training steps (T max ); -By solving the problem based on the set regression order (m) and the maximum number of training steps (T max ), during the training step, based on the training data (D T ) iteratively calculates the linear autoregressive mathematical model (M R ) is one or more parameters (θ).

10. The method according to any one of the preceding claims, characterized in that Set the linear autoregressive mathematical model (M R ) includes corresponding parameters (θ') calculated during the training step based on Ul ) and the linear autoregressive mathematical model (M R ) and calculate one or more parameters (θ” Ul ), to tune the linear autoregressive model (M R ) of one or more parameters (θ Ul ).

11. The method according to any one of the preceding claims, characterized in that The method includes the step of performing a first checking process to check the computational performance of the mathematical model (108).

12. The method according to claim 11, characterized in that The first checking process (108) includes: - comparing (108a) said first detection data (D) acquired during a predefined inspection period S1 ) and the predicted data (D P ); - calculating (108b) an error function (E), the error function indicating the S1 ) and the predicted data (D P ) is a difference between the predicted values ​​included in ; - If the error function (E) exceeds the threshold error value (E TH ), then update (108c) the autoregressive mathematical model (M R ).

13. The method according to any one of the preceding claims, characterized in that The method includes the step of performing a second checking process to check the electrical energy consumption predicted by the mathematical model (109).

14. The method according to claim 13, characterized in that The second checking process (109) includes: - Processing (109a) the calculated prediction data (D P ) to calculate a prediction function (P), the prediction function indicating a predicted trend of the electrical energy consumption in the power grid; - If the prediction function (P) takes a value higher than the maximum confidence value (P max ) or below the minimum confidence value (P min ), then an alarm signal (109b) is generated.

15. A computer program, stored or storable in a storage medium, characterized in that The computer program comprises software instructions for implementing the method (100) according to one or more of the preceding claims.

16. A computerized device, characterized in that The computerized device comprises data processing resources configured to execute software instructions to implement the method (100) according to one or more of claims 1 to 14.

17. The computerized device of claim 16, wherein: The computerized device is an intelligent electronic device (5) for a power distribution network (1).

Images