In June this year, the University of Cambridge’s Energy Policy Research Group featured Spectral in their working paper reviewing digitalisation in the energy sector. Spectral was selected as one of 40 innovative start-ups worldwide. The paper covered Artificial Intelligence, Machine Learning, Deep Learning and Blockchain applications in the energy sector with the goal of helping energy regulators calibrate their support for new business models.
Spectral featured in the section of the working paper that focussed on companies with digital solutions which utilise blockchain. The paper primarily focussed upon Spectral’s STELLAR for automated negotiation and settlement of energy and flexibility trading. This article seeks to expand that coverage to touch upon the other themes of the working paper, specifically Spectral’s implementation of Artificial Intelligence (AI) and Machine Learning.
Spectral’s BRIGHTER is equipped to help real estate owners gain greater oversight over their complete portfolio and to transform their buildings into smart, sustainable assets that are “fit for future”. Building management systems (BMS) are a critical component to manage energy demand, but often they are limited by manual interventions or static rule-based control systems. Artificial intelligence can significantly improve the efficiency and effectiveness of building management systems.
BRIGHTER consists of a number of modules which can either function as separate entities or be combined into one integrated solution to transform buildings. At its core, BRIGHTER’s dynamic building control module enables real-time, predictive optimisation of buildings to significantly improve energy efficiency and comfort levels.
The forecasting of electricity consumption plays a crucial role since it provides the basis for decision-making in asset planning and operation. In order to acquire reliable medium-range and long-range predictions, a similar-day approach for load forecasting is implemented within BRIGHTER, taking into account a number of varying factors. As explained by Iglesias and Kastner, “buildings—or buildings’ energy behaviors—are usually represented as time-based profiles or patterns to cluster” . Consequently, by combining the recognised patterns in the given time series with relevant exogenous data (such as weather), BRIGHTER takes this method one step further and is able to accurately predict electricity load by assigning every day to the cluster of the most similar historical days. This technique can also be referred to load profile clustering.
Furthermore, gas and electricity consumption forecasts are utilised as regressors for a Multi Layer Perception. This calculates the potential savings of each building for the remainder of the year, using the actual values compared against the generated forecast .
Apart from performing specific forecasts for individual buildings, data analytics within Spectral’s BRIGHTER give real estate owners greater oversight over their complete asset portfolio. The hourly electricity consumption forecast as a total across a set of buildings is used as a basis for investigating potential earnings from participation in the day-ahead electricity market. In the day-ahead market, electricity is traded one day before the actual delivery, so the bids from buyers and sellers must be known at least 24 hours in advance. Using BRIGHTER’s predictions of the hourly load using total consumption forecast, real estate owners can optimise savings by buying electricity in the wholesale market instead of via fixed contracts.
The consumption model forecasts time-series data based on an additive model where non-linear trends are fitted with different seasonalities and custom regressors . BRIGHTER’s implementation takes a vast number of parameters into account to maximise the model’s accuracy, including (but not limited to) historical day-ahead prices, weather data, similar day profiles, holidays and trends.
Another AI element within BRIGHTER is the digital twin. This is a digital replica of every element in a building comprised of blocks representing physical assets, people, places, systems and devices. This method enables us to represent any kind of building embedded with rich information about its assets and the relationships between them.
Using a neural-network framework, the digital twin fits a complex physical representation of the building and underlying systems on diverse (real-time) data inputs, such as electricity/gas/heat consumption, Building Management System (BMS) parameters, weather conditions and environmental sensor data (e.g. temperature, humidity, CO2, VOCs, occupancy), and “learns” the characteristics of the building. In this way, the building’s thermal, electrical and comfort characteristics are captured in the model, with continuous improvement as new data comes in. The digital twin model can be used to simulate the outcomes of a range of control strategies, which in turn enables the selection of the most optimal control strategy to maximize energy efficiency while ensuring high comfort levels.
The digital twin serves as a real-time system of record, offering dynamic insights on how the building is performing and drastically reducing decision-making complexity. Additionally, this accurate copy enables actions to be simulated that might otherwise be costly, risky or time-consuming to test without threatening the stability of the actual ecosystem. AI is the basis for testing control strategies into the future by fully simulating all the surrounding conditions and knock-on effects of actions, thereby enabling the optimal control strategy to be devised and tested (without any impact upon the actual building during the test phase) before deploying at the building.
Statistical machine learning is used with BRIGHTER to simulate return on investment for installing energy storage or distributed energy resources (e.g. solar panels). The business case for peak shaving in relation to the customer’s electricity connection is simulated for the battery installation. Participation in the frequency containment reserve can also be simulated. A mathematical model is built (step 1 of any statistical machine learning process) and then the historical data is used to simulate real-world conditions (step 2 whereby the AI “learns” in order to find the optimal fit – in this case, the optimised return on investment and optimal configuration of equipment, e.g. battery and inverter).
The energy transition is yielding significant changes in energy generation and consumption patterns. An increase in decentralised energy generation from renewable resources with an intrinsically variable nature results in a less predictable supply of electricity. Simultaneously, increased consumption is resulting from the electrification of transport, heating etc. and the introduction of new technologies. More consumers are becoming producers themselves, often referred to as prosumers. As a consequence, the electricity network must be able to effectively manage two-way traffic.
STELLAR is an advanced energy management platform which enables fast and easy deployment of high performance smart-grids. It allows for the optimum control of both small scale and large scale systems, providing key services to all stakeholders, from behind-the-meter customers to grid operators and utilities. The incorporation of smart algorithms and artificial intelligence within the platform significantly improves its accuracy.
Spectral deployed STELLAR on a large-scale implementation with the leading Dutch green energy provider, Greenchoice, at the Hartel II wind park. This 24 MW wind farm located by the Hartelkanaal at Rotterdam was integrated with a 10MW battery system, participating in the Frequency Containment Reserve (FCR) market. The combined system also provided portfolio optimisation services for Greenchoice. The first priority of the STELLAR implementation was to ensure that the combined operation of the wind turbines and battery system never exceeded the limits of the 24 MVA grid connection. The system enabled local peak power management via wind curtailment and active control of the battery’s charge and discharge behaviour to optimise FCR market participation, while at the same time increasing gains.
In cases where curtailment took place, the energy control system had to simulate the potential power output of the wind turbines in order to accurately calculate the savings that resulted. After extended testing, the power curve used in the first phase was replaced by a Long Short-Term Memory Recurrent Neural Network (LSTM – RNN) which performed excellently, with a measured accuracy exceeding 97.1% across all turbines. The unique property of the LSTM network to “remember” past states makes this model a success: it does not only process single data points, but also entire sequences of data.
STELLAR incorporates an imbalance price forecast. Since wind curtailment takes place at moments when feed-in prices are not favorable (e.g. negative prices), minimising the losses from wind energy generation depends on how accurately the energy control system can predict the drop in prices. These prices can be quite erratic, as they are driven by multiple highly volatile time-series and their correlations, causing simple models to fail to generate accurate forecasts. After experimenting with different techniques, Spectral developed a recurrent neural network with a custom-built attention mechanism (typically used for natural language processing), to map the driving time series and their correlations, and used this information to generate the final imbalance price forecast. Fine-tuning the network on the given time series problem produced excellent results, further maximising revenues from the combined battery and wind assets by passive balancing / imbalance price optimisation.
Artificial Intelligence allows for the performance of specific tasks quickly and effectively, relying on pattern recognition and inference. Within the subset of AI known as machine learning, a wide variety of statistical models and algorithms are used to achieve this outcome. Currently, AI is impacting a wide cross-section of industries worldwide. Applied appropriately, AI has vast potential to accelerate the global energy transition.
Spectral is already realising the potential of AI across several applications, deployed as part of several active implementations. Our product roadmap includes the continued incorporation of AI-based solutions and we expect AI to form a pivotal part of our evolving software platforms.
© Spectral 2019
1. Iglesias F, Kastner W. Analysis of Similarity Measures in Times Series Clustering for the Discovery of Building Energy Patterns. Energies [Internet]. 2013;6(2):579-597. Available from: https://www.mdpi.com/1996-1073/6/2/579
2. Kain N. Understanding of Multilayer perceptron (MLP) [Internet]. Medium. 2018. Available from: https://medium.com/@AI_with_Kain/understanding-of-multilayer-perceptron-mlp-8f179c4a135f
3. Taylor S, Letham B. Forecasting at scale. PeerJ Preprints 5:e3190v2. 2017. Available from: https://peerj.com/preprints/3190/