Source: COMMUNITY ENERGY LABS, LLC submitted to
ADVANCED GRID INTERACTIVE LOAD EFFICIENCY (AGILE)
Sponsoring Institution
National Institute of Food and Agriculture
Project Status
COMPLETE
Funding Source
SMALL BUSINESS GRANT
Reporting Frequency
Annual
Accession No.
1028482
Grant No.
2022-33530-37065
Cumulative Award Amt.
$131,500.00
Proposal No.
2022-01267
Multistate No.
(N/A)
Project Start Date
Jul 1, 2022
Project End Date
Apr 30, 2023
Grant Year
2022
Program Code
[8.6]- Rural & Community Development
Project Director
Barham, T. A.
Recipient Organization
COMMUNITY ENERGY LABS, LLC
401 NE 19TH AVE
PORTLAND,OR 972324800
Performing Department
N/A
Non Technical Summary
Policies nationwide are targeting rapid adoption of grid-integrated building strategies,time-varying electricity prices, building electrification, and resiliency. But advocates worry thattop-down innovation leaves smaller and more vulnerable communities behind. This projectbrings affordable grid connectivity to cornerstones ofrural communities: small and mid-sizedcommercial buildings (SMSCB) that have been under-represented in grid integration projectsto date. Many rural communities have older building stock and are at the end of a powerdistribution networks - meaning energy burden, power quality and outage duration tend to beworse forrural customers. Outages impact commercial building systems such as refrigerationthat are key to reducing food spoilage and contamination in agricultural production. Byshaping energy use to better match localrenewable output,reduce energy costs and demandcharges the innovation proposed in this SBIR lowers commercial electrical bills for end usersand extends the time that critical power backup can reliably provide power during outages anddisturbances. CEL's innovation is (1) to define a model architecture that can bereplicated--without significant customization --in other buildings and (2) to verify and optimizethat model architecture using real-world data and use cases in SMSCB and rural communities.This project addresses USDA's objective to to improve the quality of life in rural America bypromoting energy efficiency,resiliency and adoption ofrenewable energy resources. CELadvances deliberate strategies to make technologies for energy and demand charge reductionmore affordable in orderto meet the needs of this significant segment of the US building stock.
Animal Health Component
50%
Research Effort Categories
Basic
(N/A)
Applied
50%
Developmental
50%
Classification

Knowledge Area (KA)Subject of Investigation (SOI)Field of Science (FOS)Percent
40453402080100%
Knowledge Area
404 - Instrumentation and Control Systems;

Subject Of Investigation
5340 - Nonfarm structures and related facilities;

Field Of Science
2080 - Mathematics and computer sciences;
Goals / Objectives
Our aim for Phase I of the project is to understand the needs and use cases that typify rural andSMSCB building energy and demand charge management. Based on those use cases we willevaluate the suitability and feasibility of MPC/RL model structures to predict and control energyand peak demand in various real-world building scenarios without relying on lots of data orextensive, expensive engineering. Our primary goal is to ensure the feasibility (thermodynamically,computationally and economically) of the proposed MPC/RL framework for integration with areal-world building controller prototype that SMSCB can afford. In order to do so we need to notonly choose well, but optimize our resulting framework to ensure that its data constraints, inputs,and outputs adequately reflect conditions and data we are likely to be able to collect in our targetmarket. By working with our TAC and project partners to collect real data from their buildingswe will be able to better tailor the most feasible model for accuracy and scalability. We will usethis real-world data to verify the physics, computation time, etc. through testing and simulationand develop a functional specification, prototype deployment roadmap, workforce training andcommercialization plan for the MPC controller in Phase II. We will have met our goal in Phase Iwhen we deliver an architecture for the model framework that:1. Can be adapted to commercial constraints;2. Is applicable to real-world conditions;3. Delivers performance, in simulation using real-world data, that meets or exceeds that ofcomparable MPC architectures for computational complexity, predictive accuracy, and modeltraining instantiation time,4. Can be implemented within CEL's user-friendly software package without requiring expertsetup, making it more accessible for a wider swath of the trades to implement on the job.Specifically, we hope to converge on an architecture that can produce accurate predictions withina 6-hour prediction horizon and for which instrumentation, labor, and data costs per building areestimated to be less than $15,000 installed.OBJECTIVE 1: UNDERSTAND USE CASES' IMPACT ON MPC/RL DESIGNCEL will be primarily responsible for the fulfillment of this objective. The CEL team is led by TanyaBarham, AGILE Project Director, with participation from TAC community members and the OregonClean Energy Cooperative to recruit and gather data from rural customer sites. To successfully adaptthe technology, we must ensure that we have the correct inputs, outputs, and physical constraints.This means understanding the data-sphere (or building and operational characteristics) for ourintended use cases. CEL's Chief Technology Officer, Jennifer Worrall, or a Full Stack MachineLearning Engineer from her team, will also examine a rough diagram of the high-level data schema,common information model, interfaces, infrastructure, and other architectural considerationsnecessary for full product integration within both a prototypical controller and the CEL tech stack.This will include basic assumptions about practical aspects of full integration such as APIs betweenmicroservices, communication patterns, data quality, and cyber security. We will then use identifiedinformation to evaluate MPC/RL architecture feasibility necessary to meet Objective 2.OBJECTIVE 2: OPTIMIZE & VERIFY MODEL STRUCTURE FEASIBILITYCEL's Chief Technology Officer, Jennifer Worrall, or a Full Stack Machine Learning Engineer fromher team, will be primarily responsible for this task with assistance on structuring the MPC/RLfeasibility criteria provided by Dr. Peter May-Ostendorp. The team will test several model structurevariations in order to understand the impact of model structure and implementation requirementson model instantiation and training time, optimizing the number of inputs and outputs needed tosuccessfully and accurately function in customer use case scenarios. We aim to verify with real-world data that our model can feasibly predict and adjust control decisions and can be developedinto a Phase II commercial prototype.
Project Methods
Task 1.1 Document customer use cases and implications for the model Tanya Barham and DanOrzech will collect common use case scenarios from TAC members, cross-check use cases withClean Energy Co-op projects' power bills, and hold technical advisory meetings to determinepractical considerations necessary to architect a resulting model that is capable of achieving usecase objectives given practical constraints. In order to draft the requirements we will consider:Common Building FactorsIdentify inputs, outputs, and constraints typical of SMSCB buildings with attention paid toHVAC systems, zone configurations, and energy consumption measurement.Use Case Specific Factors• Consider zones with specialized utilization and associated requirements such as auditoriums,corridors, outdoor spaces, offices, and classrooms.• Understand the needs for integrating schedules (class, sports, extracurricular, or special events)and other disturbances (weather, power outages).Regulatory/Best Practice Factors• Review compliance implications for building performance such as ASHRAE Standard 90.1and Guideline 14 to ensure commercial use will accommodate forward-looking operationaland reporting compatibility with emerging standards, requirements, and business models.• Review and characterize additional (non-thermodynamic) HVAC requirements (like air-exchange rates) with special attention paid to potential post-pandemic requirements.Commercial FeasibilitySpecify high-level data schema, common information model, interfaces, additional sensing andcommunication infrastructure, and other architectural considerations necessary for full productintegration included in a control platform prototype.Task 1.2 Define modeling framework assumptions and data requirements CEL will de inemodeling assumptions and data requirements for the relevant use cases. Barham and Dr. Ostendorpwill identify common sets of binary, analog, and multi-state data inputs, outputs, and constraintparameters required for each of the inputs in the proposed reference model frameworks. In particular,we will analyze model sensitivity to observable and unobservable parameters in prototypicalcommercial building types (in Western climate zones). We have conducted a literature review andpreliminary inventory of relevant models. We will cross-check our work with a parameter databasethat our team developed for DE-SC0021831 Easy Does it - Automated Model Input for BuildingMPC noting typical and alternate sources for data collection and qualitative attributes of each source,such as time-cost to collect, reliability and typical ranges for the specific parameters required.Task 1.3 Document architecture, software integration requirements and evaluation criteria forfeasibility Jennifer Worrall and a CEL Full Stack Machine Learning Engineer will review workproducts from Tasks 1.1 and 1.2 to ensure that any model framework evaluated can be integrated inCEL's commercial software platform. CEL's engineering team will also suggest strategies forcleaning noisy and missing data such as forward padding, linear interpolation, or averaging andoutline means for reducing the potential impact that such data deficiencies and correction strategiesmight have on model training and accuracy. CEL's engineering team will provide Barham, Orzech,and TAC stakeholders with data collection requirements for collecting data used to evaluate andtailor the proposed model architectures.Task 1.4 Gather, audit, and clean data for MPC/RL evaluation CEL will gather, audit, and cleandata for model evaluation. CEL will gather data from selected building sites representative of morethan thirty days of nominal building operation each. Alternatively, CEL will use synthetic data fromrepresentative simulated buildings. Once CEL engineering outlines the proposed architectures anddata collection requirements (Task 1.3) with input from CEL stakeholders (Task 1.2), Barham willcoordinate with project partners to complete the steps below:Tanya Barham, TAC members and Oregon Clean Energy Cooperative will identify test sites from acohort of representative field demonstration partners. Barham will ensure data authorizations arecomplete and updated and data security protocols are followed. She will lead recruitment andselection of 3-5 buildings that are characteristic of our target market. Barham will coordinatescheduling, data collection, and troubleshooting trips to field sites.Jennifer Worrall will define the common information model and data tagging schema to be deployedusing open standards such as ASHRAE 223P/Brick if possible. She will set up a lightweight andsecure database, data transfer protocol, or secure interfaces for data exchange. Barham and partnerswill identify data input values that are infeasible or impractical to collect and, where needed, generatesynthetic data for models. Where necessary, a CEL field engineer will schedule and instrument fieldsites with additional sensors and equipment and verify sensor and API connectivity.A CEL Full Stack Engineer will perform data sampling and quality checks to ensure the data meetsmodeling specifications. They will identify whether any data sampling errors require correction strategiesas outlined in Task 1.3 and correct underlying causes of noisy and missing data with the field engineer.Task 2.1 Finalize architecture, product development plan and roadmap for Phase II prototypeBarham, along with Dr. Ostendorp and Jennifer Worrall, will evaluate methodology, outcomes,and final selection to verify we choose the optimal model structure. Based on use cases and dataanalysis from test sites, CEL will analyze feasibility of model integration in terms of accuracy;compute time; models' ability to run on modern, commercial computing platforms; level of ITinfrastructure, technical support, and maintenance each architecture will require at scale; andincremental cost the framework would add to time-costs borne by the end user relative to othersolutions, including doing nothing.At the conclusion of this task, CEL will select and document the appropriate MPC/RL andoptimization model to fit the use cases and data for the SMSCB segment. CEL will provide anM&V plan for Phase II including hardware requirements, computational speed, and prediction/control accuracy and identify bills of materials, supplies, and timeframes for development. Thiswill serve as a blueprint for CEL's prototype. It will also enable CEL or other research partners inPhase II to compare and assess the Phase II prototype design against existing modeling and controltechniques (such as PID, rules-based control or white-box MPC) deployed in current commercialbuildings technology using BOPTEST.Task 2.2 TABA & Commercialization Activities: Workforce Integration & Training OregonClean Power Cooperative will coordinate with SMSCB service providers such as solar installersand commercial mechanical and electrical contractors to understand how the proposed innovationcould fit within their existing or future service delivery models.

Progress 07/01/22 to 04/30/23

Outputs
Target Audience:Our customers are superintendents of business, COOs, building owners, facility managers, and maintenance directors for school districts, tribes, and municipalities that typically have more than 50,000 square feet under management. Commercial operations like dairies, food processing facilities, and other agricultural post-processing would also benefit from AGILE technology for their building energy management. Most of these entities have tight staffing and engage mechanical contractors through service agreements; we see this secondary reseller market as equally important. We have attempted to solve these problems by picking a market segment that we know well and that is large enough to allow us to generate profitable growth. This market also has many commonalities in building systems (e.g., packaged rooftop HVAC), budget cycles, and business processes that allow us to quickly build a core technical product and delivery model that can be adapted unaltered to other buyers in the market and then altered to fit adjacent markets once our balance sheet allows room for expansion. At scale, the flexible coordination of operational electrical end loads has the potential to provide 1393 PJ (1.32 quads) of energy efficiency and 200 GW of peak energy flexibility to the grid, a benefit of $15 billion per year to US ratepayers. But getting there means we must provide viable means for public, low-income SMSCBs to adopt grid-interactive efficient building (GEB) technologies with advanced controls that can respond to grid signals and time-varying energy prices. According to DOE's 2021 National Roadmap for GEB, "By reducing and shifting the timing of electricity consumption, GEBs could decrease CO2 emissions by 80 million tons per year by 2030, or 6% of total power sector CO2 emissions. That is more than the annual emissions of 50 medium-sized coal plants." Schools, tribal governments, and public agencies are an important subsector of SMSCBs, making up nearly a third of total US commercial floorspace. They have specific-enough characteristics for replicability of the algorithms and large enough footprints to scale and advance the state of the art. We chose this segment because the 13,720 schools in California, Oregon, and Washington are facing stringent building energy performance penalties by 2024-2026 (such as Washington HB1257 and utility time of use rates). We want to build a solution that will allow customers with limited resources to quickly and affordably improve compliance and reduce energy burden. Addressing this underserved market will be key to putting MPC within the reach of SMSCBs and achieving national and community-level energy goals by 2030. Ultimately, for an average user like a 100,000 square-foot high school, a one-time install fee of $10,000 and ongoing $.06/square foot subscription for CEL'stechnology would pay for itself in months while keeping occupants comfortable. Changes/Problems: Nothing Reported What opportunities for training and professional development has the project provided?Project Director Tanya Barham approached the rural districts included in this pilot through multiple pathways including: Utility and HVAC training programs; labor unions; and rural economic development non-profits focused on job creation in the trades - specifically HVAC, solar installation and electrical. She met with leaders from: Tacoma Power, Electric power Research Institute, Energy Northwest, PNGC Cooperative, Blachly-Lane Electric Cooperative, IBEW, Oregon Solar and Storage Association, South Sound Solar, Southern California Edison who all introduced her to trade allies and contractors that would be interested in an eventual commercial product developed through this grant. The initial 10 schools sites identified and participation in this Phase I award engaged through those channels. How have the results been disseminated to communities of interest?Throughout the duration of the Phase I Tanya Barham and consultant Dan Orzech from Oregon Clean Power Co-Op approached an additional 14 rural K12 schools districts, public library districts, community colleges and other community building portfolio owners in rural communities. At least half of these have applied through CEL's Web site http://srs.communityenergy.tech/ to pilot any technology prototype developed and deployed in Phase II. Using their and Phase I participants input CEL also performed analysis of commercialization steps at each stage of customer development for a Phase II commercial product pipeline. Namely those steps were: pre-sale, onboarding, adoption, account growth, and customer renewal. For each of these stages, we identified tasks and processes related to a customer achieving their desired goals with a particular focus on what might be required to minimize customer friction at each stage. Our technology simplifies the development of data collection and energy control models compared to other state-of-the-art approaches. It does NOT require expertise in model development; however, our conclusion in Phase I was that there are areas of complexity at each step in the above customer development stages that still require support tools to assist in the sales, onboarding, and renewal process known as customer success. Based on those findings--and after reviewing the nature of the complexity and the needs of our sales channel and our customers--we won a $100,000 grant from Business Oregon to develop "how to" videos for both our service/support personnel and for prospective customers. These support tools will smooth customer acquisition and onboarding and are central to supporting the CEL's go-to-market strategy in Phase II of this USDA project. What do you plan to do during the next reporting period to accomplish the goals? Nothing Reported

Impacts
What was accomplished under these goals? OBJECTIVE 1: UNDERSTAND USE CASES' IMPACT ON MPC/RL DESIGN Major Activities Completed The first major activity was customer use case documentation, which sought to understand the needs of rural and SMSCB buildings in relation to energy and comfort management. This led to our defining of modelling framework assumptions and the subsequent detailing of architecture, software integration needs, and evaluation criteria for MPC/RL models. As part of our investigation, we spoke to over 60 stakeholders in various roles at suburban and rural community buildings and K12 schools. This process helped us to formulate four key use cases: reducing utility demand charges, managing time of use energy consumption, resilience, and occupant comfort. We further carried out comparative studies on four algorithms - SINDY(c), Neuromancer, GnuRL, and Hybrid MPC - assessing their commercial viability based on the input data required. Data Collected: Data collection was carried out in 10 schools across 3 different climate zones in Western USA, with a specific focus on multiple thermal zones in different buildings. The collected data covered a variety of metrics, ranging from HVAC zones to the types and brands of HVAC equipment used, including their manufacturing years. We connected to the equipment's controls via APIs to collect telemetry data. This data was then meticulously audited and cleaned for MPC/RL evaluation. Summary Statistics and Discussion of Results: A significant revelation from the customer interviews was that use cases 1 and 2 held the most concern for both rural and other customers. All interviewed buyers identified energy and demand charges as significant cost contributors. The fourth use case - Occupant Comfort - was also a critical concern among potential buyers. Analysis of schools revealed classrooms constituted the largest controllable space in school buildings (30%), followed by offices (7%), and libraries (3%). Comparison of real-life data with the data required for the MPC/RL models concluded that the most suitable algorithms based on data availability were Hybrid MPC, Sindy(c), Neuromancer, and GnuRL. Key Outcomes or Other Accomplishments Realized: A key outcome was our in-depth understanding of the needs and use cases that typify energy and comfort management in rural and SMSCB buildings. Our data collection and analysis led to the development of a software backend capable of hosting any of the evaluated model frameworks on CEL's commercial software platform. We also built several APIs for efficient communication, data storage, and retrieval, with data compliance in accordance with the BRICK schema. These APIs bridge CEL's data ontology and the equipment at customer sites, facilitating model-agnostic data use. A significant accomplishment was the development of a 10-step process for data readiness, overcoming issues of missing or irregular data. OBJECTIVE 2: OPTIMIZE & VERIFY MODEL STRUCTURE FEASIBILITY Major activities completed / experiments conducted; During the second objective, we assessed the practicality and personalized adaptation of the MPC/RL models in real-world scenarios. A comprehensive feasibility analysis was conducted to determine how effectively each algorithm could be integrated into a real-world software production environment. We selected and documented the best-suited models for deployment in Phase II, considering factors such as hardware requirements, computational speed, and prediction/control accuracy. This allowed us to outline bills of materials, supplies, and development timeframes. We utilized data from objective 1 and compared the algorithms using metrics like accuracy, training compute time, runtime compute time, IT infrastructure level, and scalability maintenance. A Measurement and Verification (M&V) protocol was designed in collaboration with theUniversity of California Berkeley Center for the Built Environmentand used to conduct an initial assessment of scalability, savings, and overall economics of some early field trials. Data collected; We collected and analyzed data from the objective 1 testing phase. The comparison of algorithms was made using data collected at the same timestamps as the first objective. Our models (Hybrid MPC, Neuromancer, Sindy(c), and GnuRL) were trained and tested using this data, while specific operational details (like training time, runtime, error rates, and data history requirements) were noted. We also performed initial measurement and verification for 4 days, comparing the conventional customer schedule and Hybrid MPC. During this period, we monitored factors like ambient outside temperature and peak load reduction. Summary statistics and discussion of results; Our findings showed that the Hybrid MPC model, despite requiring the longest training time, resulted in the most accurate predictions, with a mean square error of 0.30. It can train in 27-30 minutes, run in under 5 minutes, and takes 75 iterations. It needs 5-10 days of history data for training. Neuromancer, with a 0.48 error, trains in 5-7 minutes, runs in 64 minutes with 64 iterations, and needs 30 days of history data. Sindy(c) has a 0.56 error, trains in under 5 minutes, runs in under 5 minutes with 1 iteration, and requires over 6 months of history data. GnuRL's parameters were not available. Among all algorithms studied, we selected Hybrid MPC as the most promising framework to commercialize and test in an expanded field study in Phase II, because it produces the highest accuracy with the least amount of historical data and trains within a reasonable time horizon. Furthermore, Hybrid MPC led to an average peak load reduction of 20% over a 4-day block while maintaining room temperatures within 2-5% of regular settings. This demonstrated superior efficiency and comfort control. Key outcomes or other accomplishments realized; Our evaluation confirmed that the Hybrid MPC model outperforms others in terms of accuracy and efficiency. The team is confident that we can define standard bills of materials, supplies, and timeframes for developing a commercial prototype using commercially available hardware, infrastructure, and computer resources from a variety of vendors. We have developed Hardware & Infrastructure requirements for onboarding a site and enabling data collection. We've also established guidelines for data collection, storage, and active testing and documented all the details for future reference. The project has created valuable Python libraries and tools, and extensions of the BRICK ontology specific to school-related buildings. We've built a database to store historical energy consumption, thermostat, demand, and weather data, which currently covers ten rural schools in four rural school districts. Additionally, we've laid out technical specifications required for functional design, delivery, and deployment, as well as a comparative assessment of the selected model. This equips us well for Phase II, where we aim to compare performance, cost, and supply chain from various vendors and establish relationships to improve the economics of our resulting product at scale.

Publications

  • Type: Journal Articles Status: Under Review Year Published: 2023 Citation: Ham, Sang woo and Kim, Donghun and Barham, Tanya and Ramseyer, Kent, The First Field Application of Mpc to K-12 Schools: Lessons-Learned and Savings Assessment. Available at SSRN: https://ssrn.com/abstract=4341934 or http://dx.doi.org/10.2139/ssrn.4341934