La electrical load prediction It has become a key component for modern electrical grids to operate safely, efficiently, and without unexpected bills. It's no longer enough to simply "monitor" how much energy is consumed: electrical systems need to anticipate generation hours, days, and even years in advance, such as pumping stationsdistribution, maintenance and new investments.
Furthermore, the massive irruption of the variable renewable energiesThe charging of electric vehicles and new consumption patterns necessitate the use of much more sophisticated models. From classic statistical methods to algorithms of machine learning, deep learning and Gaussian processesLoad forecasting has gone from being a relatively simple calculation to a leading field of data analytics.
What is load prediction and why is it so important?
The prediction or load forecast It involves estimating the future electricity demand of a network or facility, whether it be a home, an entire building, an industrial area, or a whole national electricity grid. The time horizon can range from some hours (short term) until several weeks, months or years (medium and long term).
Having a good forecast allows operators to decide which power plants should come online, when to do so, and at what capacityas well as defining the operating regime of the transmission and distribution networks. If the estimate is poor, problems may arise. cost overruns (because they had to buy emergency energy at very high prices) or stability issues on the internet, as in the tension control.
In the context of smart grids, load prediction combines historical data, real-time information, grid status, and external variables such as weather, calendar, and economic activityThis way it better reflects reality and facilitates a more flexible and efficient operation.
For utility companies, good load forecasting directly influences the asset planning, maintenance and financeAs many data managers in the sector point out, when demand is poorly estimated and last-minute purchases are necessary, costs skyrocket and competitiveness is reduced.
Key variables in load forecasting
The behavior of electricity demand is conditioned by many variables. Some are relatively easy to quantify and others are more unpredictable, but all influence the quality of load prediction.
Among the most important factors are the meteorological conditionsTemperature, humidity, wind, and solar radiation have a direct impact on both consumption (air conditioning, heating, cooling) and renewable energy production (wind, solar). Therefore, the load on a grid is treated as a highly random time seriesespecially in systems with a lot of renewable energy generation.
They also have a significant influence on seasonality and the calendarEnergy consumption varies between summer and winter, as does consumption on a weekday and a holiday. Load series typically exhibit daily, weekly, and annual patterns that forecasting models must be able to capture and utilize.
There are also factors linked to the Demographics and economic activityPopulation density, customer type, industrial development, or changes in consumption habits. The more detailed the segmentation (by region, city, neighborhood, or user type), the more necessary it is to have robust data integration processes and quality control over all sources used.
Finally, the growing dependence on external data (third-party meteorology, socioeconomic indicators, etc.) necessitates evaluating their reliability. A low-quality meteorological dataset can ruin a well-designed model, therefore the management of information sources It is an essential part of the forecasting project.
Classical and modern methods for demand forecasting
From a technical point of view, electric charge is modeled as a time seriesThat is, a sequence of observations ordered in time. Historically, the first load forecasting methods relied on the classical statisticsalthough today they coexist with much more advanced models.
Among the traditional models are the autoregressive (AR) models, those of Moving average (MA) and their integrated combinations (ARMA, ARIMA). These models assume that the future of the series can be described in terms of its past values ​​and certain noise terms, seeking to minimize error metrics such as Mean percentage absolute error (MAPE) or the mean squared error.
With the expansion of computing power and data availability, the machine learning (ML) It has gained significant prominence, especially in short-term forecasting. Algorithms such as support vector machines (SVMs)Gaussian processes, neural networks of all types, or hybrid combinations of statistical and ML models.
SVMs for regression are often complemented with ARIMA models and use nonlinear and cyclic kernels which better capture daily or seasonal patterns. Gaussian processes adapted to time series also provide a probabilistic estimation of the load, providing not only an expected value, but also a measure of uncertainty.
Regarding neural networks, multilayer perceptrons gave way to more complex architectures such as LSTM and GRU (recurrent) networks, deep convolutional networks (CNN) and other deep learning models capable of automatically extracting relevant features, at the cost of a greater need for data and computational power.
Probabilistic models and load interpretability
A very interesting line of work in load prediction combines predictive power with interpretabilityIn other words, the goal is not only to make an accurate estimate, but also to understand What patterns and relationships explain those predictions?This is done using probabilistic models that explicitly describe the underlying structure of the data.
These models allow one to obtain, naturally, a predictive variance associated with each estimate, which indicates the degree to which we can trust the result. At the same time, its parameters usually have a clear interpretation (for example, covariance matrices that reflect correlations between times of day or between regions).
A prime example is the approach of load profilingMachine learning algorithms are used to identify profiles or groups of behavior in demand series. These profiles serve to simplify the understanding of the system and to train it. specialized models by type of profileimproving performance.
Instead of treating prediction and profiling as separate steps (first grouping, then fitting regressors), some models integrate both phases into a single probabilistic formulation. A representative case is the clusterwise linear model (CWLM), based on a mixture of Gaussian regressions but augmented with a linear regression component in each cluster.
In the CWLM, the expectation maximization (EM) algorithm simultaneously adjusts the clustering and regression parametersTherefore, the groups found are directly optimized for the prediction task. This allows for capturing profiles such as seasonal patterns (summer/winter) or differences between weekdays and weekends, while also improving accuracy.
Multi-tasking prediction and advanced Gaussian processes
Another powerful idea in this area is to treat load prediction as a multitasking problemFor example, estimating the entire next 24 hours or the loads of several neighboring regions at once. In these cases, the outputs are clearly correlated with each otherAnd leveraging those relationships can substantially improve performance.
The Multi-tasking Gaussian processes They are a natural extension of classical Gaussian processes in which the covariance between tasks is explicitly modeled. This approach provides not only more accurate predictions, but also inter-task covariance matrices interpretable that allow you to visualize which times of day or which areas are most related.
A recent formulation is the so-called Cool-MTGP (conditional one-output likelihood multi-task Gaussian process). Its strategy consists of associating each task with a single-output Gaussian process conditioned on the previous tasks, reducing the complexity and the number of parameters with respect to standard multi-task formulations.
The Cool-MTGP avoids typical problems such as having to estimate large inter-task covariance matrices using low-rank approximations, which usually introduces additional hyperparameters. Instead, it maintains the interpretability of covariances with a more controlled computational cost.
Applied to a real-world 24-hour forecasting scenario with aggregated data from a regional operator (such as ISO New England), this type of multi-tasking Gaussian process can outperform other benchmark models and offer a very accurate estimation of time correlations, without the need for overly complex cross-validations.
Load forecasting according to the REBT: buildings and low voltage
Beyond the major networks, the low voltage load forecast This is the starting point for sizing conductor cross-sections, protective devices (circuit breakers, miniature circuit breakers), and transformers in building and premises installations. In Spain, all of this is regulated by the Low Voltage Electrotechnical Regulation (REBT), especially by ITC-BT-10.
ITC-BT-10 indicates how to estimate the expected power in different types: buildings mainly intended for housing, commercial or office buildings, industrial buildings, concentrations of industries and parking lots with electric vehicle charging.
In general terms, the computing power is obtained by adding the contributions of each zone or use (dwellings, general services, premises, garages, EV charging) and applying, where appropriate, simultaneity coefficients which take into account that not all loads will operate at maximum power simultaneously.
Furthermore, the REBT differentiates between single-phase and three-phase suppliesDistribution companies are obliged to guarantee the operation of any single-phase receiver up to 5.750 W at 230 V, the usual maximum single-phase power being 14.490 W (63 A at 230 V)Above that value, the normal practice is to switch to a three-phase supply to avoid unbalancing the network.
Load forecasting also has effects on the hired potency by the user. Although the planned power is mainly used to size the installation, the customer can contract a lower value, depending on their actual usage habits, always within the limits set by the IGA and the distributor.
Simultaneity coefficient and total power calculation
In electrical design, an installation is almost never sized by directly adding components. all nominal powers of the equipment. If this were done, the infrastructure would be oversized in most cases, since it is very unlikely that all receivers will operate at full load at the same time.
To solve it, you use the simultaneity coefficient (CS)This represents the fraction of the installed capacity that is expected to be demanded simultaneously. Mathematically, it is the ratio between the maximum foreseeable simultaneous capacity and the total installed capacity of the entire installation or part of it.
In collective installations (apartment buildings, shopping centers, industrial parks, EV charging stations, etc.), CS is especially relevant because the statistical effect of many users makes the probability of perfect match low. As the number of consuming units increases, the simultaneity coefficient tends to decrease.
For most practical cases, the CS is not calculated from scratch with general formulas, but rather... values ​​tabulated in the regulations or in criteria established by the distributors themselves, based on experience and historical studies. When a specific table does not exist, professional experience, analysis of actual load curves, or other energy auditing methods are used.
Once the appropriate CS for a group of loads has been determined, the simultaneous calculation power is obtained very directly: simply multiply the total installed power multiplied by the simultaneity coefficient, which provides the design power on which conductors and protections are sized.
Housing, general services, premises and garages
In buildings primarily intended for housing, the total planned power PT It is composed of the sum of four main blocks: housing complex (PV), general services (PSG), Commercial premises and offices (PLC) y garages (PG)Each one has its own rules of calculation.
In the case of homes, the power to be planned is linked to the assigned intensity of the IGAFor new constructions, the minimum power is 5.750 W at 230 V (basic electrification level), rising to 9.200 W for homes with high electrification (more surface area, electric climate control, automation, many points of use, EV charging in single-family homes, etc.).
The total number of dwellings in the building is grouped by applying a specific simultaneity coefficient depending on the number of dwellings and their level of electrification. ITC-BT-10 provides a table where, for example, for 15 dwellings a value of 11,9 is considered, representing how many equivalent dwellings are estimated to be able to simultaneously demand its maximum power.
The general services load is calculated by adding the power ratings of elevators, lifting devices, pressure groups, heating and cooling plants, entrance lighting, stairwells and common areas, without applying simultaneity (CS = 1). However, in motors and discharge lamps, it is necessary to consider correction coefficients according to ITC-BT-47 and ITC-BT-44.
For commercial premises and offices, the standard establishes a minimum of 100 W/m² per floor and a minimum power per premises of 3.450 W at 230 V, with a simultaneity coefficient of 1. If the actual calculation of the planned load is higher, the actual one is taken; if it is lower, the minimum regulatory amount is applied to guarantee sufficient capacity.
In the case of garages, ventilation is taken into account: 10 W/m² for natural ventilation y 20 W/m² for forced ventilationalways with a minimum of 3.450 W at 230 V and CS = 1. In addition, if there is electric vehicle charging, the specific power of the charging points following the rules of ITC-BT-52.
Special receivers: discharge motors and lamps
Some equipment, by its nature, requires a load forecast higher than its nameplate rating. This occurs mainly with... electric motors (due to the starting peak) and with the discharge lamps (due to the apparent power they handle).
In discharge lamps (fluorescent, mercury vapor, sodium vapor, metal halide, etc.), the minimum design load in VA is taken as 1,8 times the power in watts of the lamps, according to ITC-BT-44. This correction seeks to better reflect the reality of the apparent power that flows through the conductors.
For motors in general, ITC-BT-47 indicates that the conductors that supply to a single engine They must be sized for 125% of the full load current, and that when they are fed several engines125% of the largest engine plus 100% of the others is taken. In terms of expected power, this is equivalent to multiplying by 1,25 the largest engine and add the rest without factoring.
Engines elevators, cranes and lifting equipment They still have a larger margin: for forecasting purposes, the nominal power is multiplied by 1,3When the exact power of an elevator motor is unknown, ITC-BT-10 provides tables with typical values ​​according to load, speed and type of device.
In mixed installations with motors and other loads, the calculated intensity or power is obtained by adding the corrected part of the motors and the expected power of the rest of the loads, applying simultaneities only when appropriate and according to the specific use of each group.
Electric Vehicle and LGA Protection Systems (SPL)
The recharge of electric vehicles (EVs) It introduces a new consumption block that could become very relevant. ITC-BT-52 establishes that the charging forecast for EVs in collective parking facilities is calculated by multiplying 3.680 W for 10% of the parking spaces built.
The resulting power is integrated with the rest of the building's loads, but a simultaneity factor This differs depending on whether a Main Supply Line Protection (MSL) system is in place. With an MSL, a CS of 0,3 is typically used; without an MSL, CS = 1 is used.
El SPL It acts as a kind of "smart ICP" (circuit breaker) on the building's main electrical panel. Instead of abruptly cutting off the supply if a threshold is exceeded, the system monitors the current, temporarily reduces the power allocated to EV chargingIt disconnects non-priority loads or regulates the load intensity to avoid overloads.
This system is composed of current meters, electronic controllers y actuators that decide which loads are disconnected or limited. Thanks to IoT technologies and the potential integration of artificial intelligence algorithms, SPLs will gain the ability to make proactive decisions based on predicted load patterns.
Although it is not always mandatory today, everything indicates that, as the fleet of electric vehicles grows, Having SPL will be almost essential to avoid over-dimensioning the infrastructure and to manage peak demand without compromising the network.
Taken as a whole, load prediction combines regulations, classic electrical design, and advanced data modeling algorithms to anticipate demand, correctly size installations, and operate networks in the safest and most economical way possible, both in large transport infrastructures and in everyday buildings.
