Resumen de Power losses estimation in low voltage smart grids
One of the European Union Targets was to replace at least 80% of all traditional energy meters with electronic smart meters by 2020. However, by the end of 2020, the European region (EU 27 including the UK) had installed no more than 150 million smart electricity meters, representing a penetration rate of 50% for smart meters. By 2026, It is expected that there will be more than 227 million smart meters in households due to the updated planning and target numbers, which will affect many European markets, including western and northern Europe. This scenario would contribute to the general purpose of building a more sustainable distribution system for the future.
A LV smart grid is characterized by the existence of an advanced metering infrastructure (AMI), which provides a bidirectional channel of communication among smart meters. The data collected by smart meters can be used to improve distribution network operations such as network balancing, demand response and energy flexibility. Low-voltage smart grids are essentially distribution networks that contain a secondary substation (SS) with a power transformer that steps down the voltage from the medium voltage (MV) side (20 kV) to the LV side (400 V). There is also a set of feeders or individual lines that supply power to residential, commercial and industrial customers. The feeders can be underground or aerial.
This thesis contributes to the field of power loss estimation and optimization in low-voltage (LV) smart grids in large-scale distribution areas. To contextualize the importance of the research, it has been necessary to explain the unbalanced nature of low voltage distribution areas where there is a huge deployment of smart meter rollout, and there is also uncertainty related to renewable energy generation. Main results of the thesis have been applied in two smart grid research projects: the national project OSIRIS (Optimización de la Supervisión Inteligente de la Red de Distribución) and the European project “Ideal Grid For All” (IDE4L).
Smart metering infrastructure allows distributor system operators (DSOs) to have detailed information about the customers energy consumption or generation. Smart meters measure the active and reactive energy consumption/generation of customers using different discrete time resolutions which range from 15–60 min. A large-scale smart meter rollout allows service providers to gain information about the energy consumed and produced by each customer in near-real time. This knowledge can be used to compute the aggregated network power losses at any given time. In this case, network power loss are calculated by means of customers' smart meters measurements, in terms of both active and reactive energy consumption, and by the energy measured by the smart meter supervisor located at the SS.
It has to be noted that the coverage of smart meters in LV distribution networks is not 100%. Some customers still have analogic meters, which means that only a monthly energy measurement is available for those customers. Hence, in some LV networks, there exist non-telemetered customers which provide data about their energy consumption/generation on a monthly basis. In these situations, there is a great uncertainty associated to the customer demand in shorter periods of time (day, hours).
Moreover, these non-telemetered customers are usually medium scale consumers who have a contractual power superior to 15 kW, consequently, their impact on network power losses is considerable. This situation complicates the total network power loss calculation because there is not accurate information about the demand of non-telemetered customers in near real time (15–60 min).
The problem of network losses estimation becomes more difficult due to the existence of not-technical losses due to electricity fraud or smart meter measurements anomalous (null or extremely high) or even because there are customers’ smart meters that can be out of service.
One of the differential keys of LV smart grids is the presence of single-phase loads and unbalanced operation, which makes it necessary to adopt a complete three-phase model of the LV distribution network to calculate the real value of the power losses. This scenario makes the process of power loss estimation a computationally intensive problem. The challenge is even greater when estimating the power losses of large-scale distribution networks, composed of thousands of SSs.
Nowadays, sustainability has become a key aspect to ensure a safe, reliable, and an affordable energy supply for the future. The decarbonization of power systems is an essential task to achieve a sustainable power supply. Hydrocarbon-based resources are limited, which means that actual consumption trends will be strained in the medium term. The use of renewable and sustainable energy sources is a mandatory path that society will have to follow. Nonetheless, the transition cannot be performed in a straightforward manner or in the absence of energy policies and mechanisms to ensure a reliable and efficient power supply. The sustainable path will involve the adequate and progressive incorporation in LV distribution networks of renewable-based distributed energy resources (DERs) – such as photovoltaic (PVs) panels, plug-in electric vehicles (PEVs) and battery energy storage (BES) devices. Moreover, the combination of PV panels and PEV and BES devices can contribute to smoothing the power demand curve through flexibility mechanisms.
In recent years, environmental concerns have led to the increasing integration of a considerable number of DER devices into LV smart grids. This fact prompts DSOs and regulators to provide the maximum energy efficiency in their networks (i.e., the smallest power loss values) and maximum sustainable energy consumption. Detailed understanding of the network’s behavior in terms of power loss and the use of electricity is necessary to achieve this energy efficiency.
However, the above scenario presents some drawbacks. The integration of DERs units (such as PV) into distribution networks can produce an increment of network power losses if the DERs units are not optimally located, coordinated, or controlled. Additionally, the network can experience technical contingencies such as cable’s overloads and nodal over-voltages or can lead to an inefficient system operation due to high energy losses or cables that exceed thermal limits. Moreover, there is a great uncertainty associated with the distributed power generation from PVs because its energy generation depend on weather conditions, including ambient temperature and solar irradiance, which are highly intermittent and fluctuating. Uncertainty is also present in some loads with stochastic behavior, such as PEV devices, which adds an uncertainty layer and makes their optimal integration more complex.
Therefore, DSOs require advanced methods to estimate power losses in unbalanced large-scale LV smart grids under uncertainty situations. Such estimations would facilitate the deployment of policies and practices that lead to a safe and efficient integration of DERs in the form of flexibility mechanisms. In this context, flexibility mechanisms are essential to achieve optimal operation conditions under extreme uncertainty. Flexibility mechanisms can be deployed to tackle the imbalance between generation and demand that results from the uncertainty that is latent in LV smart grids.
These flexibility mechanisms are based on modifying the normal power consumption (for the demand side) or power generation (for the generation side), according to a flexibility scheduling at the request of the network operator. In the case of the demand side, it can involve moving the energy consumption from one period of time to another that is more convenient, with the reward of an economic incentive. This mechanism is known as “demand shift” and it requires the customer to actively participate in a demand response (DR) program for fulfillment the network operator requests. Another demand-side flexibility mechanism is load curtailment, in which a series of maximum power consumption set points is defined for the customer during a certain period. If the customer fulfills the flexibility schedule, they are rewarded with an economic incentive.
On the generation side, a flexibility mechanism also exists, which involves modifying the active and reactive power injected to the grid. The flexibility mechanism entails either the renewable energy curtailment or the regulation of the reactive power injection. However, the determination of the flexibility schedule for DER sources is very complex under uncertainty conditions such as the stochastic and intermittent PV generation. Hence, relying on the flexibility provision in such devices can be potentially risky if the flexibility is required to keep the system within normal and safe conditions. In summary, DSOs face the challenge of managing network losses over large geographical areas where there are hundreds of SSs and thousands of feeders, with multiple customers and an ever-increasing presence of renewable DERs. Power loss estimation is thus paramount to improve network energy efficiency in the context of the European Union energy policies. This situation is complicated by the unbalanced operation of those networks and the presence of uncertainty. To address these challenges, this thesis focuses on the following objectives: 1. Power loss estimation in unbalanced LV smart grids under uncertainty 2. Power loss estimation in unbalanced LV smart grids in large areas with a presence of DERs 3. Flexibility scheduling for power loss minimization in unbalanced smart grids under uncertainty The mentioned objectives are achieved by taking advantage of smart metering infrastructures, machine and deep learning models and mathematical programming techniques which allows DSOs to reduce their total power losses within the distribution network. This approach entails using flexibility mechanisms to operate the distribution network optimally and enhance the load management and DG expansion planning.
To address the first objective, a methodology for power loss estimation in smart grids under uncertainty is formulated in this thesis. Load demand from the customers’ smart meters was gathered through the AMI communications infrastructure. The existence of non-telemetered customers and the missing load demand data were considered through a top-down approach. Missing load demand data were estimated through an NLP optimization process, which minimizes the quadratic error between the predicted load consumption measurement and the reference value. The reference value was obtained by using historical consumption data providing an energy consumption tendency as well as weekly and hourly energy consumption patterns. Those patterns decompose the monthly energy consumption of the non-telemetered customers and provide an estimation of the weekly, daily energy consumption. The patterns were modeled according to the type of customer (residential, commercial, or industrial), the area in which they are located, the contractual power and the presence of DERs. The Non-Linear Programming (NLP) optimization problem had constraints associated with the daily energy consumption measured by the smart meter supervisor located at the SS.
Once the hourly load consumption of the non-telemetered customers was estimated, an intra-hour high-resolution load demand profile was synthetically generated. This load profile represents the trajectory of the customer’s load demand across one hour. This is done for both customers with smart meters and those with analogic meters. The hourly-load consumption was modeled as a Markov process, in which for every stochastic realization, the total technical power losses of the network is calculated from the unbalanced power flow. Then, the evolution of technical power losses with the demand was statistically modeled through a linear model to yield a linearized power loss model.
To fulfill the second objective, a comprehensive and reliable deep learning-based power loss estimation model for unbalanced LV smart grids with DERs in a large area was proposed. First, a set of network operation characteristics was described to represent both the architecture and the operational behavior of the LV distribution networks operating in the large-scale area. Then a feature extraction process based on principal components analysis (PCA) was carried out to obtain a reduced set of features that captured most of the variability of the feeder's data. This was treated as a preliminary data preprocessing step to reduce the dimensionality of the original dataset of network features to a smaller set of relevant features. Following that, a K-means++ clustering process was applied to obtain the representative feeders that were the closest feeders to the centroids of the clusters. Therefore, the selection of the representative feeders relied on the relevant features defined in the previous step.
The next step was the use of a deep learning-based power loss model to infer the power loss level of the representative feeder, based on the demand and generation conditions. The chosen model was a deep neural network regressor model. Therefore, the power loss model was represented by a regressor supervised model, formulated as an active power loss estimator for the whole LV distribution area under study. To train the model, we synthetically created a set of demand-generation scenarios for the representative feeders; their corresponding power losses in unbalanced conditions were calculated using the three-phase power flow model depicted in this thesis. The set of synthetically generated scenarios was composed by aggregated patterns of demand and generation for each representative pattern, along with the real power loss values calculated from a three-phase power model.
The power loss inference for the rest of the feeders that belonged to the distribution area was performed based on the Euclidean distance between each feeder and its corresponding representative feeder. The learning process of the model followed the K-fold procedure, which also yielded the optimal combination of the model's hyper-parameters. The model customized with the optimal hyper-parameters was then fed the demand and generation data from the metering infrastructure of the representative feeders, scaled up to the number of feeders belonging to the feeder's existing cluster. The power loss model was set in production in a real utility LV distribution area in Madrid, Spain. It exhibited good performance as a power loss estimator and improved on the traditional approach by solving the computationally expensive power flow equations for each individual feeder or calculating time-consuming.
The proposed methodology thus appears to be a potential real-time operation tool to improve the energy efficiency in large distribution areas with a high penetration of distributed resources. The proposed method provides a timely tool for DSOs to infer the power loss levels in large-scale distribution areas using big data and deep learning technologies. The power loss model was formulated using Python as the programming language as well as the Sci-kit Learn and Tensor Flow libraries. Training scenarios were generated using the Statsmodels Python library.
To address the third objective, we proposed a robust flexibility scheduling optimization model for one day ahead, for unbalanced LV smart grids. The model considers customers’ demand as a flexibility mechanism. The objective is to minimize the total flexibility cost required to operate the unbalanced network at a minimum power loss level while keeping the voltage and current magnitude within statutory limits. The objective involves both flexibility commitment decisions and flexibility dispatch decisions – that is, the amount of flexibility to be provided by the customer. The optimization model includes as uncertain variables the PV injections and the PEV charging demand. Both uncertain variables are characterized by polyhedral uncertainty sets configured with an uncertainty budget. The model also includes the three-phase network equations and the phases statutory limits. The resulting optimization problem is a mixed-integer non-linear (MINLP) optimization problem within a min-max optimization framework, which is solved using an iterative-strategy using a cutting plane solver.
The problem was formulated in the Pyomo-RoModel framework (Python) and was solved with a Couenne solver. A comparison was provided with a stochastic programming model to minimize the expected value of the flexibility cost across a set of scenarios. The scenarios were built using a SARIMA model, yielding a forecast prediction set of values and a probability associated with each value for the uncertain variables, at each time step for the day ahead. A third comparison was based on the optimal power flow, treating the uncertain variables as parameters, and using the expected value for the day ahead. Numerical simulations were conducted on a modified IEEE European LV test feeder. The results indicated efficient solutions in terms of computational cost and robustness. The starting situation of the test feeder involved overvoltage and overloading due to excessive PV injection into the network. The application of the proposed robust flexibility scheduling model alleviated those contingencies and minimized the power losses of the network.
According to the objectives identified earlier, the main contributions of this thesis are the following: • Power loss estimation in unbalanced LV smart grids under uncertainty conditions o An optimization-based procedure to estimate load consumption of non-telemetered customers.
o A Markov chain-based process to estimate intra-hour load demand for data having a low resolution and for non-telemetered customers or customers who display incorrect measurements.
• Power loss estimation in unbalanced LV smart grids in large-scale areas with a presence of DERs o A data mining approach to reduce a high-dimensionality dataset in smart grids to yield a reduced set of relevant features.
o A clustering process to obtain representative feeders within a large-scale distribution area of smart grids.
o A deep learning-based power loss estimator for large-scale LV smart grids. The method is formulated as a deep neural network that uses as input features the power load demand and power generation of a set of representative feeders. The model gives, as output, the power losses of the whole area.
• Flexibility scheduling for power loss minimization in unbalanced smart grids under uncertainty o A robust optimization model for the flexibility scheduling optimization model for unbalanced smart grids with distributed resources, such as PV panels and PEV devices.
