Posted on

Anya’s Status Report 04/26

  • Things done:
  • Changed the ML model flow as follows
  • Training → Save as .pkl → Backend loads .pkl → Predict → Dashboard displays

  • Trained the models offline to predict solar generation and grid prices, then saved the trained models as .pkl files. This separated the training phase from the backend operations. In the backend server, implemented code to load the .pkl files at startup.

    Predictions are passed to the frontend dashboard through FastAPI routes, where they are visualized alongside real-time sensor data. This setup made the dashboard faster and more reliable, since model loading happens once at server start instead of during each user request.

    This method avoids long inference delays and made it easier to update models by replacing the .pkl files when retraining was needed.

    Did some user testing to see whether people would actually use the product

User Satisfaction Metrics

  • UI Navigation:
    Measures how easily users were able to navigate and interact with the SmartWatt dashboard during their first use.

  • Feature Usability:
    How intuitive and useful users found key SmartWatt features
    (Average score: 8.0/10)

  • Schedule Understanding:
    Evaluate whether users understood the optimized appliance schedules generated by the backend system.

Progress

Almost done, now need to just prep setup for the demo

Things to Complete Next Week

  • Video
  • Work with Erika and Maya to finish getting everything working together in the model house
  • Poster
  • Demo
  • Final Report
Posted on

Team’s Status Report 04/26


Risks

During demo testing, router issues on Maya’s laptop temporarily disrupted Raspberry Pi connectivity. We are switching to a new device router for stability. This can be a network setup challenge since we havent fully tested communication over this new network.

Changes

We changed the setup to use a standalone router for the Pi. the Pi connects to Wi-Fi network using the router.

Progress

  • Unit Testing 

    • Tested device control APIs

    • Evaluated forecasting model accuracy (MAPE/RMSE) on solar and price data.

    • User Testing & User satisfaction

  • Frontend-Backend Integration Progress:

    • Connected backend schedule outputs to frontend dashboard for live visualization.


Unit Tests Conducted

  • Device Control Test: Verified API calls successfully switched smart devices ON/OFF at the correct time by looking at system logs
  • Optimization Solver Test: Validated LP outputs feasible schedules under varying constraints, edge cases include flat/spiky data for prices and solar output
  •  Forecasting Model Test: Evaluated regression predictions against known solar generation and pricing data.
  • UI Functionality Test: Confirmed correct display of real-time power data, scheduled actions, and user overrides.
  • Front-End API Connectivity Test: Confirmed that React frontend could successfully call FastAPI endpoints and receive responses.

 

Overall System Tests:

  • End-to-End Scheduling Test: Simulated a full daily cycle from sensing → predicting → optimizing → scheduling → controlling devices.

  • Stress Testing: Input flat load profiles and sudden random spikes to test optimization stability and system robustness.

  • Latency and Responsiveness Test: Measured UI response time under background optimization tasks.

  • End-to-End API Test (Chatbot): Tested communication between chatbot UI and backend, ensuring user input is captured and responses are sent back

Findings and Design Changes

  • Edge Case Handling: Stress testing exposed situations when no feasible schedule existed (overlapping tight constraints); we added fallback rules.

  • API Logging Addition: Initial debugging was slow without detailed API logs. We added full request/response logging to trace failures and latency bottlenecks.

  • Shift to Asynchronous Processing: Originally, synchronous optimization caused UI freezes. Moving to background task execution using FastAPI improved responsiveness.

  • Prediction Model Tuning: Retrained regressions models with better feature scaling and added time-of-day features to improve accuracy in volatile solar conditions.

Posted on

Team’s Status Report for 04/19

Risks

  • Model Deployment Latency: ML model loading during user requests (e.g., price forecasting) caused site latency and crashes.  The models were being loaded inside the FastAPI route handlers at runtime during each user request. Every time a user accessed the forecast route, the model had to be deserialized from disk which introduced delays Will move the model loading logic to the top-level script (the FastAPI startup event or global cache)
  • Ensuring all the wiring and hardware components have enough physical space in the demo house to be properly integrated and mounted.

Changes

ML workflow is being restructured:

  1. Preload models at startup using global caching to ensure they are loaded once and reused.

  2. Introduce a prediction cache, where forecast results are computed periodically (e.g., every 30 minutes) and stored in memory or a lightweight JSON/SQLite file.

Progress

  • Device actuation and control complete
  • Device scheduling complete
  • Testing and Verification of Optimization subsystem complete
  • Routing for backend : Separated all core functionalities into clean routes (/, /optimize, /analysis, /history, etc.) with a responsive dashboard UI.
  • Working on integrating the python backend with typescript front-end.
Posted on

Anya’s Status Report for 04/19

Work Accomplished

Integrated the devices (Big LED, PWM Fans) to the dashboard.

  • Implemented real-time device control through the dashboard with “Turn On/Off” capability.

  • Optimized the calls for device state and history queries to stay under 500ms latency.

  • Added schedule-based control logic using persistent JSON storage and periodic polling, enabling automated device operation aligned with user preferences.

  • Tested the optimization algorithm. Measured the latency, with the full scheduling computation consistently completing within 500ms to 1.5s. This ensures that users receive near-instant recommendations
  • Evaluated how well the algorithm responds to price shifts by injecting synthetic price spikes in the mid-afternoon. The optimizer avoided those periods unless solar generation was high, confirming that the algorithm appropriately weights cost considerations. In solar alignment tests, we simulated peak solar output at 11 AM–2 PM and observed the optimizer shifting device usage accordingly, validating the integration of renewable energy forecasts into the objective function.
  • Verified solver stability and objective sensitivity to weights which include grid price and solar output.

Progress

  • The Software – Embedded Layer actuation is done.
  • I faced issues with ML model integration—currently, the models for forecasting grid prices and solar output are being called during route execution, which causes high latency and even site crashes (too long to respond) in some cases. I still need to figure out how to load the trained models separately and keep them persistently available to the app (running inference in a background job queue) rather than reloading them with each request. I am slightly behind schedule on that integration.

Next Steps

  • Add More Devices to Dashboard:

    • Extend the devices list to include the other speaker and motors. On the dashboard will create a filter to view  power sensor mapping for each new device for usage tracking and optimization.

Refactor current ML routes to load models once during app startup instead of per-request.Use a tool to preload models and serve fast predictions.

Knowledge and Tools

  • FastAPI & Jinja2 Templating: I had to learn FastAPI for building a responsive backend, and how to integrate it with Jinja2 for rendering dynamic HTML dashboards.

  • Home Assistant API Integration: I learned how to query real-time sensor states, parse device schedules, and handle authentication with Home Assistant’s REST API.

  • Linear Programming with PuLP: I explored energy cost minimization and scheduling using linear programming, learning how to define decision variables, objective functions, and constraints.

  • Data Visualization (Chart.js, ECharts): I gained familiarity with frontend charting libraries to visually present sensor data, solar forecasts, and results in user friendly way.

  • Model Deployment (joblib, XGBoost, Scikit-learn): I had to train and serialize ML models for solar and pricing forecasts, then load them into the backend

Posted on

Team’s Status Report for 04/12

Risks

  • We are training LSTM/CNN models based off of data polled from sensors.  Retraining the model every request (if the cache fails) can cause lag

  • Home Assistant API latency spikes can delay forecasts. Plotting charts via matplotlib adds overhead (200–300ms)

  • Caching logic failing silently if timestamps are misaligned

  • Infeasible LP results due to overlapping constraints or underforecasted solar

  • LP objective may optimize cost but ignore user experience (e.g., clustering devices)

  • Forecast errors can propagate into LP, yielding poor schedules

Changes

The initial approach used an LSTM network to forecast solar generation based on time-series data from sensor.maya_solar_panel. Main challenge was high latency due to recurrent nature and sequential computation and overfitting on short term trends rather than daily cycles. Transitioned to a 1D Convolutional NN. CNNs capture short-term temporal correlations ( sunrise to peak curve) better.

Progress

  • Successfully polled data from all sensors into backend

  • CNN model produces realistic solar forecasts aligned with production hours

  • Performed testing on API response times

Verification/Validation

  • Used curl requests to profile API timing in the terminal while polling data from sensors

  • Unit tested LP constraints to ensure feasibility and logic correctness

  • Compared cached vs. uncached outputs for CNN model for correctness and stability

  • Compared forecast vs. actual solar power to assess prediction alignment
  • To validate the 15% energy savings, we’ll compare historical baseline energy usage (from fixed schedules) against SmartWatt’s optimized schedules using the same input conditions. Using sensor data (from sensor.maya_fan_power, sensor.maya_solar_panel), we’ll simulate both scenarios over a 1–2 week period and calculate daily energy costs. The optimized case will use CNN-based solar forecasts and LP scheduling. If the average cost reduction exceeds 15% compared to the baseline, the savings target is met.
  • To verify the robust assembly of the house, I will carry the house around and apply reasonable force to ensure it will remain intact during the demo.
  • To validate the integration of the hardware components with the house, I will perform a fit test with the interior components. I will ensure that all feedthrough holes are wide enough and that the components can be neatly and securely placed inside their respective rooms.
  • For validation, we plan to conduct user testing sessions where participants interact with the model house and dashboard to simulate their household energy consumption. This will help us evaluate whether SmartWatt creates a coherent and immersive experience. We will analyze user feedback to determine if the system effectively communicates the intended purpose of the project, the scaled energy usage and passage of time, and feels intuitive to interact with.

 

Posted on

Anya’s Status Report for 04/12

Accomplishments

  • Integrated real-time sensor polling and displayed it on the dashboard

  •  Developed a responsive FastAPI + HTMX dashboard to display live sensor values

  • Implemented Chart.js charts that auto-refresh every few seconds without full page reloads

  •  Latency about 10s for the API calls, page refreshes every 30s

Progress

Reduced latency by introducing in-memory caching for the a CNN solar forecast endpoint

  •  Forecast results are now cached and refreshed every 10 minutes, eliminating redundant model retraining on every request

  • Slightly behind progress, need to integrate all the charts in the same app with routing

  • Verification

Performed basic verification of the CNN forecasting model to ensure it meets the intended design requirements for solar prediction accuracy.

  • Input Feature Sanity Checks:
    Verified that time-series input includes correct power, sin_hour, and cos_hour features, resampled to 15-minute intervals.
    Resampling was validated via debug logging (len(df), delta distributions).

  • Window + Horizon Coverage:
    Model was trained on sliding windows (e.g., 48 steps = 12 hours) with a 24-step forecast horizon. These parameters were confirmed to provide full diurnal coverage and enough context for trend learning.

  • Envelope Alignment Verification:
    The daylight envelope was implemented and centered around peak solar generation (13:00–14:00) to prevent unrealistic output at midnight. Forecasts before and after masking were compared to confirm correction.

  • Polling Interval Consistency:
    The system pulls data using HA/api/history/period endpoint. I confirmed that sensors such as sensor.maya_solar_panel, sensor.maya_fan_power, sensor.maya_fan_voltage, and sensor.maya_fan_current are polled with sufficient granularity — typically every 5–10 minutes.

  • Each response is parsed to validate:

    • state is numeric and finite

    • last_changed timestamps are increasing and consistent

    • There are no large gaps in time , > 1 hour)

  • Resampling Debug Logs:
    Added internal logging to capture the number of raw and resampled datapoints. This ensures the model always receives enough sequential history for the sliding window.
    Used interpolation (.interpolate()) on the resampled time-series to ensure continuity in data, even if occasional values are missing from Home Assistant.

  • latency tests via time curl http://localhost:5051/api/endpoint to make sure the dashboard refresh updates meet design and usecase requirements

Next Steps

  • Test LP behavior under forecast uncertainty

    • Run LP with low vs. high solar predictions

    • Assess how schedules shift and whether they violate any constraints

  • Stress-test LP under edge-case constraints

    • Zero solar availability

    • Multiple high-power devices requiring overlap

    • Very short vs. very long runtime constraints

Test LP reactivity to live data

  • Feed in updated solar forecast every hour

  • Ensure LP returns updated schedules quickly (<2 sec)

 

Posted on

Team’s Status Report for 03/29

Risks

  • Backend performance under real-time load
    With multiple device schedules and real-time data polling, system lag or crashes may occur if not optimized.
  • Home Assistant integration 
    Integration with HA might face compatibility or API syncing issues depending on setup.
  • Edge-case handling for devices
    Unresponsive or offline devices could cause unexpected failures if not properly handled

Changes

  • System currently loads price data from CSV files, need to change it to connect to a SQLite database, for storing device state, energy usage, and forecasts is in place
  • Add some more loads such as speakers for an audio component in the model house

Progress

  • Frontend and backend frameworks set up
  • User preferences taken into the consideration when performing constrained optimization
  • Model house for demo set up
  • Sensors integrated into Home Assistant
Posted on

Anya’s Status Report for 03/29

Work Accomplished

  • Implemented conversational AI interface allowing users to query the system about energy optimization at their home.
  • Created suggestion system that guides users with prompt examples for better engagement with the AI assistant
  • The optimization system was enhanced with a user preferences framework that allows the user to select custom timeslots in which to run devices.
  • The system incorporates earliest start times and latest end times for each scheduled device.
  • Then designed a new constraint-based optimization that respects user preferences while maximizing energy savings
  • A three-tier priority system was implemented:
    • Low Priority: Maximizes energy savings with flexible timing (priority weight: 0.3)
    • Medium Priority: Balances energy savings with user-preferred times (priority weight: 0.5)
    • High Priority: Strictly adheres to user time preferences (priority weight: 0.8)
  • Priority settings directly influence how the optimization algorithm weighs time constraints against cost savings

Progress

Frontend and backend up and ready.

Right now, I need to implement the API routes that will actually trigger device control — switching devices on or off based on user input or automated schedules. Slightly behind schedule with regards to the GANT chart.

These routes will act as the bridge between the UI actions and the actuation layer. Once the endpoints are set up and mapped to the appropriate device control logic, the system will be able to execute real actions, completing the loop from user interaction to physical outcome.

Next Steps

  • Tie each route to the code that interacts with the device (ESP 32GPIO pins/ Home Assistant API).
  • Then test with real devices to validate actual switching.
  • Refine ML or linear programming algorithms that decide when to turn devices on/off.
  • Incorporate feedback loops from usage data.
Posted on

Anya’s Status Report 03/22

Work Accomplished:

  • Did some testing of a LSTM neural network architecture for time-series energy consumption prediction
  • Tracked self-consumption metrics for solar utilization, peak reduction and cost savings % according to a baseline.
  • The original cost is based on the predicted load with original device schedule. The optimized cost uses the new schedules with shifted loads

Mock backend for schedules : (I need to parse the results of optimization algo and figure out whats a good way to put that into device schedules). This is what the frontend should look like

Current display of predicted vs optimzied power. Need to scale this up accordingly once the power sensors are connected to Home Assistant and download the data from there.

Progress: I would say I am a little behind because a) the power sensor data is not logged in, need to do the integration of the backend with Home Assistant to feed that data, look at transience to analyze average power consumption b) The grid pricing data is fed in through Nordpool which has an integration  with Home Assistant, but for now I am training the LSTM by downloading a csv from Home Assistant rather than having it dynamically via an API (API calls are expensive)


Challenges and Next Steps:

  • Improve load shifting algorithm with better device priority handling [figure out a way to parse info about optimized loads from an aggregate to a device level]
  • Also right now, all devices have the same priority.  Need to assign weights based on device priorities (like running a fridge vs fan, fridge is way more important)

  • Use these weights in the objective function to favor high-priority loads when minimizing cost or peak demand.

  • Add detailed logging for performance metrics to validate optimization results

  • Integration between HA and backend would be just creating a docker container and adding that to HA vs API requests
Posted on

Team’s Status Report for 03/15

Risks:

Gradient Step Update Issue:

  • In SGD and L-BFGS optimization algorithms, when forecasted energy demand closely matches actual load, the gradient of the cost function is near-zero.
  • This results in very small updates to optimization variables, reducing adaptability and making it harder for the model to improve scheduling decisions.
  • Mitigation: Introduce regularization techniques, perturbations in forecast data (white Gaussian noise)

Misinterpreting Transients as Long-Term Load Changes

Short-term power spikes (when a fan starts) can be mistaken for sustained high energy demand. The optimizer may overcompensate by reducing battery discharge or increasing grid imports unnecessarily.

Changes:

No major changes happened this week. We are thinking of adding a user tracking functionality to improve future recommendations based on past actions to make the system more interactive.

Progress:

  • Optimization Model Debugging: Refined solver selection and fixed SOC linearity issues.
  • Created CAD model of the model house
  • Algorithm Development: Partial success with SGD & L-BFGS, but linear programming constraints need debugging.
  • Recommendation Engine: Classifies insights by priority and impact, with user tracking integration.
  • Frontend: Built an interactive energy visualization dashboard.
    Pending:
  • Hardware Integration: Waiting for power sensor polling to complete before full backend integration.
  • Building the model house in which to house all of our electrical components.