Skip to content

Latest commit

 

History

History
138 lines (112 loc) · 6.77 KB

TODO.md

File metadata and controls

138 lines (112 loc) · 6.77 KB

TODO

Steps In progress

  • Is taking the mean of an average ok? Should I try with Energy content (that doesn't solve the probably above with electricalProductionLoadDiff)

Battery charging problem

  • Fix problems with predictions: test different models:

    • Check that I have the correct shape. When using this function in the example repo: [1, _.size(featureDescriptions)]

      • Error: Error when checking : expected dense_Dense1_input to have shape [null,12] but got array with shape [12,1].

    • Make sure training and test data is lined up correctly (features line up with targets)
    • Try LOTS of epochs (50)
    • Plot Predicted vs. Actual in Tensorflow examples to compare with this one
    • multiLayerPerceptronRegressionModel1Hidden and multiLayerPerceptronRegressionModel2Hidden
    • Use and optimize linearRegressionModel layers using DSS
    • Understand data normalization better and make sure I'm doing it correctly. XGBoost worked well, but also works well with unnormalized data, which means I may not be normalizing it correctly (gracefully handles un-normalized or missing data, while being accurate and fast to train)
    • Use different splits of the data
    • Check how validation is done in the current script
    • Split battery model into it's own store

  • Fix productionLoadDiff: It should be Total Renewable Power Output - Total Electrical Load Served (won't affect calculations yet)

  • Add button that retrains model

  • Save trained model to localstorage: https://js.tensorflow.org/api/0.14.1/#tf.Model.save

Table fixes

  • Only add column headers and units when rendering table (don't store those with the data)
  • Calculate new combinedColumnHeaderOrder as a computed function in store based on excluded columns
  • Add Column Selector to combined table
  • Filter columns in combined table
  • Store column filters in localstorage
  • Make sure table updates on homer or appliance change
  • Consolidate tables
    • Comment out appliance and HOMER table
    • Put battery stats in Battery Energy Content tab
  • Have hover tooltip in column indicator. Click turns it on or off
  • Turn on or off blocks of columns in with the legend
  • Clean up styling in column selector indicator
  • Color code dropdown column selector menu (and decrease line height)
  • Add Hover on table

TODO:

  • Check that HOMER and Appliance file has required fields (including battery and pvSystem)
  • Indicator that we are currently calculating
  • Deploy Sentry to production
  • Unmet load sums by hour, by day. Try 100% stacked.
  • A new bar chart: (histogram by hour)
    • New appliance load
    • Excess production
  • Render appliance summary stats
  • Min State of chart is fragile. Find a more robust way of finding the floor
  • Add hour offset input(?)
  • Parse Table dates
    • Show chart by datetime instead of hour of year (toggle between hour and datetime)
    • Page through datetime chart by day, week, month
  • Show column stats (max, min, avg, std) for HOMER and combined tables

TODO: Long term

  • Reference Lines: http://recharts.org/en-US/examples/LineChartWithReferenceLines
  • Load other HOMER and appliance files in the background
  • Snake case all header names. I can have a lookup to display them nicely (generic_1_k_wh_lead_acid_asm_energy_content)
  • Measure battery capacity in the field and compare with HOMER models. Quantify
  • Dropdown of selects for keeping and removing table colums

Done

  • Speed up calcs by not merging tables - just keep calculated columns in their own data structure. Pass all of those data structures into React Virtualized to display them but not merge them. That will save ~1second per update

  • Replace combinedTable with calculatedColumns

  • Consolidate all util functions into a single file

  • Remove activeHomer.keyOrder and just pass the table. Search for

    • activeHomer.tableData and tableData
    • activeHomer.keyOrder and keyOrder

  • [?] Pass in both and calculatedColumns and activeHomer into summaryStats()

  • Color headers in homer, appliance and combined

  • Render kw_factor from appliance (and color it) in CombinedTable

  • Check that combined table matches deployed production app

  • Color Combined Table columns to show calculated columns

  • Met and unmet demand didn't change in percent no matter what appliance. It was driven by system type.

    • This is working correctly

  • Round kw_factor, round homer values

  • Parse all dates to ISO 8606, reformat for display in tables

  • Create a better error for when a file isn't found (currently it's found in the csv parsing step and things the rows aren't correct)

  • Sentry should only trigger on production

  • Make sure all tables update when changing data:

    • Appliance table: change appliance
    • HOMER table: change HOMER
    • CombinedTable: modelInputs, HOMER file change, Appliance file change
    • Unmet loads update with modelInput, HOMER and Appliance changes

  • Add Grid to top of every table, add is loader spinner

  • Write 'renders without crashing' tests for all components

  • Rename battery from 'Generic 1 kWh Lead Acid [ASM]' to Battery through

Battery Charging Characterization

  • Pass in battery data and break into training and test data. Refer to https://towardsdatascience.com/train-test-split-and-cross-validation-in-python-80b61beca4b6
    • trainFeaturesArray,
    • trainTargetArray,
    • testFeaturesArray,
    • testTargetArray
  • Convert to normalized tensors
  • Create model of battery State of Charge based on electricalProductionLoadDiff and prevBatterySOC (tensorflow)
  • Lay out battery page with
    • training progress
    • predicted vs actual values
    • Sorted Weights
    • Score
  • Create derived column productionLoadDiff = PV Power Output - Total Electrical Load Served. Instead of pulling from only PV Power Output, it needs to also take into account diesel or other generation. I should probably create an additional variable. So create 2:
  • Create new page for battery characterization
  • Save trained model to store
  • Generate predicted vs actual values
  • Chart predicted vs actual

Battery charge calculation notes

Goal: Calculate Battery SOC for every hour. Use the spreadsheet values without new loads applied to generate the SoC based on loads we know.

The SOC depends on:

  • Charge/Discharge rate. These depend on
    • Excess
    • Demand
  • Previous hour's SOC

The predictive model may just want to calculate Charge Rate and Discharge Rate (or Max Charge Rate and Max Discharge Rate)

  • Removing
    • Available Capacity: Because it depends on Excess Electrical, SoC and Energy Content for the given row.