Category Archives: HEMS Project

Washing on sunshine

As regular readers will recall I’ve recently updated my battery controls with solar predication. For some years now my battery storage has automatically charged during sunshine and later discharged in the evening, and over the last year I added the HEMS to manage buying electricity including for the battery when the electricity cost was cheapest as there’s little solar in the winter, but now the HEMS has the ability to automatically adjust how much power is bought from the grid depending on how much output is expected from the solar panels in order to deliver a charged battery by the end of the day when the electricity price rises significantly. The battery charging can thus swing from all solar to all grid and all points in between entirely automatically based on the solar forecast.

Predicted solar output for March 16th to 18th as of the evening of March 15th.

Now however I’ve also added the capability to adjust when the wet goods – dishwasher and washing machine – run as a function of expected solar panel output.

As with the battery controls, I don’t attempt to match generation half hour by half hour with device operation because of uncertainty around how precisely a solar forecast from the evening before will match actual solar production up to twelve hours later. Instead my solar algorithm extracts the number of hours of significant solar production and the earliest start time of that production. For the battery that information is used to modify the number of hours of battery charging to be bought from the grid and the end of the window during which those hours of charging may be bought. Now I’ve updated the controls for the wet goods similarly.

The existing wet good controls look to start the machine at the time when the electricity cost for the cycle is cheapest within a time window set by the user, so the user defines the earliest and latest acceptable start times and the algorithm finds the cheapest start time within that window. The updated wet goods controls assess the number of hours of solar charging available and if both (i) the window exceeds a certain size and (ii) the user’s time window includes the solar window then the user window is narrowed to start at the start of the solar window. The resulting start time between the start of the solar window and the end of the user window may not be absolutely the cheapest grid cost but my assumption is that the solar contribution (which could be up to 100%) will in practice make this the cheapest grid cost as any power needed from the grid will be at a relatively good price. If this start time in the solar window does not reflect absolutely the cheapest grid cost then a second start time may also be identified which is the cheapest grid cost.

HEMS schedule with start times for wet goods within the solar window.

In the above example the start times are both within the solar window and the cheapest energy price, so no cheaper alternative is also offered. Start times within the solar window tend to be in the afternoon as the grid energy costs are lower. This also improves the probability that the battery may briefly discharge if the total load exceeds the solar output. However in the above example the energy price is so low that the battery is also force-charging (it’s dark green) so any surplus demand will come from the grid.

HEMS schedule with start times both inside and outside the solar window

In the above example the cheapest time to run each cycle from the grid is at night, although given the availability of solar during the day then any small saving in grid costs at night is very likely wiped out by the ability to run some (if not all) of the cycle during the day from the solar surplus rather than the grid. Both start times are available – the absolutely lowest grid cost and the lowest grid cost during the solar window.

It’s probably also worth mentioning the implications of running the appliance alongside the battery status being different colours:

  • Blue battery and state of charge being at or over 80% – appliance is prioritised over battery allowing battery to discharge to meet appliance needs as required – least likely to draw anything from grid (but likely highest cost to draw from grid)
  • Light green battery or (blue battery and state of charge being below 80%) – appliance is prioritised over battery charging alone (so battery may not discharge to support appliance) – a little more likely to draw anything from grid (but likely mid cost to draw from grid)
  • Dark green battery – battery charging and appliance are equal priority – most likely to draw something from grid (but likely cheapest cost to draw from grid, even to the point of being paid to draw from grid at times)

Tuning up for the performance

Last night rather unexpectedly my solcast solar irradiance data tuned itself. I use this data to predict the output of my solar panels and adjust what I buy from the grid in response. I had expected that tuning would happen eventually, but my understanding was that two month’s data was required, not the two week’s data that I had so far supplied.

Prior to tuning my predictions had looked like this..

solcast predictions before tuning

Although the system is clearly predicting output to a reasonable degree of accuracy, there are two obvious issues:

  1. The orientation of the array from due south seems a little off, as my array starts and ends generation earlier than the prediction suggesting the the orientation should be slightly more easterly in the model.
  2. The peak at a sustained 4 kW is overly optimistic. The panels can generate 4 kW according to my monitoring, but only relatively briefly and certainly not for more than an hour in March.

Nevertheless, the overall identification of better and worse days is clearly working.

solcast predictions after tuning

However after tuning both issues have been resolved. The measured and predicted curves match very closely. Note that the prior predictions have been updated by the tuning process, so March 26th which is included in both the images looks subtly different. Note also that my control is based on forecasts from the evening before, not the ‘1 hour ahead forecasts’ illustrated above. You may also observe that there is a discrepancy in the morning of April first, where the measured data is zero while the estimated actuals and forecasts are quite healthy, which arises as a result of a temporary failure of the immersun server from where the data is taken.

The ability to do tuning requires an upload of data. I upload generation data continuously at 5-minute intervals (the shortest allowed interval) which may explain the early availability of the tuned results. The script that I use to achieve this takes data indirectly from my ImmerSUN and is modified from this script. I achieved a 99% correlation which is pretty good. Subsequently it seems that the tuning takes place automatically each day.

Tonic solar, a little light music

As winter turned to spring my thoughts for my HEMS turned to thinking about how to adjust the operation of the HEMS in managing battery charging to account for anticipated solar production. Previously the HEMS was configured to buy a preset number of hours of charge from the grid each day, typically overnight when the power tends to be cheapest. Through 2019 as the seasons changed I periodically adjusted this figure to create headroom to store charge from the solar panels later in the day. However I would like to make this adjustment automatically day by day.

Late last year I came across solcast a website that predicts the output of solar panels. The user creates an account, describes their PV array (location, capacity and orientation), and can then download predictions via API.

Solar prediction for the next three days

The orange line shows the predicted output for the next few days, with the light grey area showing the confidence interval from 10 to 90%. As a prediction there’s a degree of uncertainty associated with the prediction as there is with a weather forecast. The 10% line shows that 1 day in 10 the output will be lower than the grey area, while the 90% line shows that 1 day in 10 then output will be higher than the grey area.

My original prediction were based on the orange line (the 50th percentile) where output was equally likely to be above or below this amount. However my risk on an incorrect prediction is not even. If I fail to buy enough power from the grid when the price is low then I risk paying up to 35p/kWh to buy when the price is high, whereas if I buy an unnecessary cheap kWh from the grid I may spend an unnecessary 5p/kWh on average. Thus I decided to take a more conservative position on risk to obtain the lowest cost position. I opted for a 20% risk, so 1 day in 5 I might underestimate my purchase from the grid reflecting the ratio of grid prices high:low. I estimated this 20th percentile assuming a normal distribution on the low side where the 20th percentile = 0.34 * 50th percentile + 0.66 * 10th percentile.

The process is as follows:

  1. download the data in a half-hourly format to be comparable to the half-hourly Agile electricity price data,
  2. calculate the 20th percentile from the 10th and 50th percentiles,
  3. count the number of half-hours in the 20th percentile data above a threshold that provides for charging the battery at full power,
  4. establish the earliest half hour when solar would charge the battery at full power,
  5. adjust the period for buying power from the grid by the period anticipated for solar charging (tomorrow a total of 7 hours is to be achieved by 6.5 hours from solar leaving 0.5 hours from the grid),
  6. adjust the end time for buying power from the grid to align with the earliest hour when the battery can charge at full power from solar.
HEMS schedule with battery behaviour modified for predicted solar generation

The result of these calculations can be seen above. One half-hour of lowest price battery charging has been identified overnight to meet the requirement for charging from the grid. Normal HEMS behaviour has also identified several other periods of grid charging at lower cost during the day, but these are not counted towards the target for buying from the grid due to the potential for a double count of solar and grid charging during the day. (In a similar manner there are multiple start times during the afternoon when the dishwasher or washing machine can be run on grid power more cheaply than the optimal overnight start time.)

No explicit command from the HEMS is required to enable proportional charging of the battery from the solar panels when there is a solar surplus as 2 of the 3 used battery operating modes (normal and charge-only) have this capability, and the third mode force charges the battery. If the battery is force charged during solar surplus then the source of the energy will of course be the solar panels, but any shortfall will be met from the grid.

modecharge behaviourdischarge behaviourcomment
NormalProportional to solar surplusProportional to shortfallUsed at high grid prices (today discharging enabled at > 8 p/kWh)
Charge onlyProportional to solar surplusNo dischargeUsed at mid grid price to save stored energy for period of higher grid price (between 5 and 8 p/kWh today)
Force chargeFull powerNo dischargeUsed at lower grid price to buy from grid (today buying at < 5p/kWh). If low price occurs at time with reasonable solar generation then use of solar output will happen automatically, with only any shortfall coming from the grid.
Battery behaviour in different operating modes

The cost of smart

Discussion elsewhere prompted me to look into what I spent on what you might term my energy smart systems relating to electricity consumption, so I thought I’d document it here.

ItemDescription CostComment
1Solar photovoltaic system (4kW)£5,500Bundled with ImmerSUN.
2Powervault battery storage (4kWh)£2,000Free installation as part of UKPN trial.
3ImmerSUN management system with monitoring.£600Estimate based on today’s pricing.
4Remote-controlled car charger.£300Modified used charger from eBay. My own software.
5Raspberry Pi items to make HEMS£200My own software.
6Wet goods automation (WIFIPLUG x 2)£70
TOTAL£8,670

Prior analysis of items #1-#4 in pre-Agile days has suggested a total of 9 years to achieve payback on this investment through use of around 85% of the generated energy. Solar panels are potentially good for over 20 years operation, although I doubt the lead-acid batteries will still be operating for anything like that long.

The combination of item #5 with my Octopus Agile dynamic smart electricity tariff has resulted in my average bought electricity price being 7.75 p/kWh in 2019, about half the UK average. I suppose that I could make the same judgements and program items manually each day, but the HEMS significantly reduces my time commitment to achieve that.

Item #6 is my most recent addition. The sophistication of the algorithm combining the Agile tariff with a simple model of the cycle of each device is such that I would never achieve such a high quality result manually. However the saving is perhaps only a three pence each day so maybe £10 per year on my Agile tariff and thus 7 years to pay for the two smart plugs.

Much of this content is thus around 7 years to payback. The HEMS is potentially much quicker, but relies on having smart systems to control such as battery storage and car charger.

The ‘Appiest Days of My Life

One of the consequences of integrating a smart home is the large number of different apps, web portals and potentially sources of APIs involved. The ones I use include:

TitleAppPortalAPIPurposeComment
BrightYNYReads and stores consumption from smart meter.No price data for my tariff due to smart meter limitations.
EveY /3NNEve’s alternative to Home for all HomeKit accessories with additional functionality for Eve’s own devices.I prefer this to Home for editing rules.
I use Eve products mostly for central heating control.
HomeY /3NNApple’s own app for the HomeKit smart home ecosystem.Need to refer to device manufacturers own apps (such as Eve or WIFIPLUG) for some configuration and data.
HEMSNYN My own web portal to view HEMS schedule and status via Apache web-server on Raspberry Pi.
MyImmersunYYY /1Control of ImmerSUN power diverter.Available API provides some measurement and status data as per main screen of the app.
PowervaultNYY /2Control of Powervault storage system.Available APIs provide some user scheduling and status capability.
OctoWatchdogY /3YYFuture cost, and historic costs and consumption (30 prior days) from Octopus (electricity supplier).APIs provided by Octopus.
App developed by an enthusiast using Octopus APIs.
Octopus’s own web portal provides historic consumption but does not pair this with cost. Monthly statements show graph of consumption and cost for each day.
WIFIPLUGYNYControl and measurements from own brand smart plugs.Plugs also appear in Home and Eve apps.
I use for dishwasher and washing machine.

Notes to table:

  1. APIs not officially released. Reverse-engineered by an enthusiast and available on line.
  2. APIs not officially released. Used as part of a sponsored trial when I first got the battery and re-used by myself with some manufacturer support.
  3. iOS only. Not available for Android.

Some of these apps have similarities:

  • Both Bright and OctoWatchdog show whole of house energy consumption (and potentially cost) derived from the smart meter. However they have differences too. A smart meter sits on two networks: (i) the Wide Area Network (WAN) via which the meter communicates with the energy supplier and (ii) the Home Area Network (HAN) which links the devices in the home (electricity meter, gas meter, CADs/IHD and gateway). Bright connects to the HAN via small piece of hardware called a Glow Stick Wi-Fi CAD and collects its own data in real time and stores its own records of energy consumption in the cloud; while OctoWatchdog involves no extra in-home hardware, and takes data a day in arrears from Octopus not storing anything in the cloud itself. Bright’s USP is the real time consumption and current day’s data (neither of which OctoWatchdog supports), while OctoWatchdog’s USP is the availability of electricity price which isn’t available from the meter.
  • Both Eve and Home interact with all devices in the whole HomeKit ecosystem. Eve is best for creating rules and has more ability to configure Eve’s own devices, while Home is best for sharing access with family members. WIFIPLUG’s app is more limited only interacting with their own devices, and thus cannot see Eve or other HomeKit devices.
  • Both MyImmersun and WIFIPLUG apps, and the Powervault portal, allow configuration of their own manufacturer devices. They all have, for example, timer capability and data logging. MyImmersun is better for giving a whole-of-home view showing solar panel output and net input to house (so provides a more comprehensive energy monitor), Powervault shows no solar panel output but does give a view of whole-of-home, while WIFIPLUG provides only a view of the energy consumption of devices plugged in to the WIFIPLUGs.

HEMS2 – improving the breed

It’s now around a year since I first started with my HEMS. Initially it was managing just the charging of my electric car around the cheapest electricity prices, but subsequently I added water heating, storage battery control and, most recently, dish washer and washing machine.

That original HEMS was built around a used Raspberry Pi that I bought cheaply from a colleague but, while its processing power was perfectly adequate for the task (which isn’t at all demanding), it did have some mechanical issues. Firstly, the mounting holes for accessory cards (known as HATs – HArdware on Top) moved between generations of Raspberry Pi and the holes in my Pi for the standoffs did not align with those in the newer accessory card. Secondly, I never managed to find a case with the space for the HAT and the mains cables that it switches. The former resulted in the electrical connector between the Pi and the HAT taking much of the mechanical load, and an occasional loss of connectivity to the HAT.

These mechanical issues are resolved by the new HEMS which uses a third generation Raspberry Pi allowing use of proper standoffs and a new case with the depth to enclose the HAT too.

Relay HAT on Raspberry Pi 3 (without standoffs fitted or insulation layer).
HEMS2 (the black box) replacing my original HEMS

As previously two cables connect to the HEMS – a black USB cable bringing 5V power and a grey multi core cable that brings a live mains feed to the relays with three switched lives returning to the adjacent junction box. The relays control my less intelligent loads – the ImmerSUN and the car charger – while other more intelligent loads are controlled via APIs via Wi-Fi and the cloud. The whole assembly continues to be mounted on a board on the side wall of the airing cupboard. When modification is required the assembly can be lifted off two screw heads and laid on the floor.

All the software on both generations of HEMS is the same except for the scripts that interact with the relays – either to set or read status. I assume that the pinouts on the Pi must have changed between generations as an identical HAT card moved from the ic0 to the ic1 bus requiring a single digit change to each interacting command.

Raspberry Pi images are licensed under the Creative Commons Attribution-Share Alike 3.0 Unported license.
Attribution: Efa at English Wikipedia

Testing Times

So, over the last few days I’ve been adding control of my wet goods (dishwasher and washing machine) to my HEMS so that it can start washing loads at the optimum time – that is the times with the lowest projected electricity cost for a typical wash cycle.

Electricity prices for 18th and 19th December.

The above image shows the prices for December 18th and 19th as downloaded by the HEMS at 16:45 on the 18th after publication of prices for the 19th. The HEMS then reviews this pricing against the need for electricity to determine when to use electricity and, in the case of the battery when to discharge the battery. I currently have the HEMS configured to provide:

  • 3 hours of car charging
  • Not more than 2 hours of water heating (but only when electricity is cheaper than gas which can also heat water)
  • 6 hours of fixed battery charging.
  • Best start time for dishwasher and washing machine.
HEMS plan for 18/12 17:00 to 19/12 17:00.

I’ve described the operation of car charger, water heating, and battery previously; the new content here is the two final columns on each side for dishwasher and washing machine. The numbers in the columns are the estimated cost of a washing cycle if started at the beginning of the corresponding half-hour. The yellow colouring reflects the selected half-hour to start the appliance i.e. the one with the lowest estimated cycle cost. There’s also the option to set a threshold above which one is not prepared to pay to wash today which has resulted in the red box. If the whole day was red then the washing cycle would be deferred for consideration the next day.

History from Dishwasher WIFIPLUG

The extract from the WIFIPLUG history for the dishwasher shows the typical operating sequence:

  1. @22:19 I turn the WIFIPLUG on to enable the dishwasher to be programmed and the cycle started. The plug for the dishwasher is inconveniently at the back of a low kitchen cupboard so this was achieved via the WIFIPLUG app.
  2. @20:20 I turn the WIFIPLUG off via the app suspending the cycle in its first moments.
  3. @02:00 the HEMS turns the plug on remotely allowing the dishwasher cycle to continue.

The equivalents for the first two actions for the washing machine can more conveniently be achieved via the push button on the WIFIPLUG itself as the plug for the washing machine is above the counter.

Electricity consumption 19/12.

The above image shows the measured power consumption from my smart meter. Almost 9 kW is being drawn at times under the action of the HEMS when the electricity price is cheapest, but also zero at times when the price is highest. The washing machine contributes to the peak spike around 02:00 when both it and the car charger are enabled. The later spiking during the peak period is the electric oven cycling on and off under control of its thermostat as the battery isn’t powerful enough to run the oven, so the excess power is drawn from the mains.

Electricity consumption and price 19/12.

The screenshot above shows the half-hourly electricity consumption and costs from Octopus. It should be noted that this is half-hourly consumption in kWh whereas the prior chart was power so, for example, an average 6.6 kW of power consumption results in 3.3 kWh of energy consumption in a half-hour. That the blue line of price is almost the inverse of the red cost columns indicates the HEMS is doing its job to shift most energy use to when the energy price is lowest, and use the battery to offset demand when the price is highest.

Overall on the 19th I paid £1.07 for 21.776 kWh of electricity including standing charge which is independent of use. That’s 86 p for 21.776 kWh without the standing charge, or a weighted average of 3.95 p/kWh. That weighted average of 3.95 p/kWh compares to a range of 1.10 to 26.24 p/kWh during the day. According to UK power the average cost of electricity in the UK is 14.37 p/kWh so I paid 27.5% of the average UK price on December 19th.

The 5 Step Program

This post describes the evolution of my HEMS code to control my dumb wet goods (dishwasher and washing machine) using smart plugs.

The program for my HEMS works as described below. For clarity I’ve emboldened the new steps associated with the control of the wet goods:

  1. Download the cost data for Agile from Octopus. The API provides 48 hours of data, but I use only 24 hours at a time. I download at 16:45 to create a schedule from 17:00 today to 17:00 tomorrow. I use two fields only: the date-time stamp and the energy price inc-VAT.
  2. Calculate cycle cost. Reverse sort the unit cost data in descending time, and combine the energy price with the load profile for each non-interruptible device (dishwasher and washing machine) to estimate the cost of running a washing cycle starting on each half hour. Add as third and fourth fields to the data file.
  3. Establish start time for each non-interruptible load. For each appliance in turn, sort the cost data in ascending cycle cost. Enable the appliance for interval with the lowest cycle cost within an acceptable time window (typically the first row), and overwrite remaining instructions from the prior day. Repeat for other appliances.
  4. Establish on times for each interruptible load. For all interruptible loads (battery, car charger and immersion heater) sort the data in ascending unit cost. Replicate the unit cost column for each interruptible load. For each load enable for the required number of half-hourly intervals within the time window set by the user, and disable for higher cost half-hourly intervals.
  5. Prepare user screens. Sort data file by ascending time, split into first and second 12-hour periods, and present as two HTML files.
Shell script to create daily schedules and user screens.

The top level scheduling script which runs automatically at 16:45 each day is a shell script which calls a series of awk scripts to: (i) calculate cycle costs for wet goods, (ii) determine start time(s) for each wet-goods appliance, (iii) determine on/off times for interruptible loads.

Awk is a pattern-matching program for processing text files. Such text files may be thought of a series of records and fields in a textual database. Awk may seem an odd choice of scripting languages, but essentially the processing of a text file of 96 time and unit pairs to create HTML files of 48 times and cost combinations is a text file processing task. Along the way as each on or off decision is made a system call to the OS is made by awk to copy an on or off script to a time-stamped script (e.g. either washingmachineon.sh or washingmachineoff.sh is copied to washingmachine_0930.sh). An internal timer called cron runs each half-hour so, for example, at 09:30 it runs all the scripts with 0930 in their names which updates the status of each controlled device.

At the time of writing we have 5 HEMS-controlled devices:

  1. Battery storage – interruptible – the only device with 3 states (normal, charge-only, force charge).
  2. Car charger (EVSE) – interruptible – on/off.
  3. Dishwasher – non-interruptible- on only.
  4. Immersion heater – interruptible – on/off.
  5. Washing machine – non-interruptible – on only.

Characterising the loads

The loads that I already control via the HEMS (that is loads which can be interrupted like battery charging, car charging and water heating) are essentially constant, that is that they draw the same power regardless of progress. The battery charging does start to tail off at particularly high states of charge, but that effect is neglected here. However the non-interruptible loads like the dishwasher and washing machine are expected to have very variable loads as the cycle continues – largely because periods of water heating demand much more power than spraying water around, spinning the drum, or pumping water out.

I thus decided that I would measure the pattern of these variable loads and use that information as part of the decision when to run the load for lowest cost. Rather than tabulate the half-hourly energy price as I do for interruptible loads, I would instead estimate the energy price to complete a load for each half hour in which I might start the cycle. This process of measuring the load pattern during a cycle I have referred to as characterisation.

I characterised the loads using a plug-in energy meter reporting kWh which I read manually every half hour during a typical wash cycle to determine the blue lines above. I then calculated the energy usage in each half hour period as per the orange line. I did all of this half-hourly as that’s the interval after which the energy price changes and correspondingly the interval at which the HEMS updates its output controls.

The measured loads confirmed my suspicions regarding variation with heat inputs providing the largest loads – twice for the dishwasher as it both heats the water at the start and heats to dry the dishes at the end, while the washing machine also heats the water at the start but spins to dry the clothes at the end (which is more energy efficient). The effect of this is that the washing machine would generally be expected to start in the cheapest half hour as that’s when it uses most energy, while the optimum time for the dishwasher is more complex to determine.

Of course the actual loads for the wash cycle will vary from the prediction for various reasons including not only the ability to select different wash cycles and options but also the smartness of the device in assessing how full it is or how dirty the utensils are.

I also decided that I would not explicitly turn these loads off from the HEMS – only turn them on. Not turning the loads off allows for some uncertainty as to the length of a cycle depending on the actual cycle selected, options selected, loading of machine etc. Not turning the outlet off also provides for the user being able to do additional loads under manual control if at home later in the day.

Internally my HEMS creates a schedule of 48 half-hourly actions each day for each load. Typically each action is either ‘on’ or ‘off’ (the fixed battery is more complex). The actions for these devices will follow a similar pattern of 48 half-hourly actions, but typically only one of those is ‘on’ and the 47 actions that would normally be ‘off’ do nothing. There existence is largely for equivalence and re-use of code, but they also ensure that the previous day’s ‘on’ command is over-written as required if the start time for the new day is different to that for the prior day.

Choosing the smart plug

One of the enhancements to my HEMS that I’ve had in mind for some time is to control wet goods – that is dishwasher and washing machine. I had previously thought that this might have to wait until I replaced my current machines with smart equivalents, but recently discovered that both existing machines recover after a power outage and continue their cycles. This creates the opportunity to put the machines on smart plugs, to manually start washing cycles, but then immediately turn the smart plugs off, and then use the smart plug to resume the cycle at the optimum time as instructed by the HEMS.

I already have two Eve Energy smart plugs as part of my smart heating system using Apple HomeKit, but these aren’t ideal for this application as I really need an exposed API to allow the HEMS to have control. However I then came across WIFIPLUG which, not only has an exposed API, but also Apple HomeKit compatibility (and indeed compatibility with other smart home ecosystems).

The WIFIPLUG (unlike the Eve Energy) also features an integrated on/off button allowing the washing cycle to be paused using this button, or via the HomeKit app, before later resuming under HEMS control when the energy cost is lowest.

The WIFIPLUG also has its own app which allows timers to be set and energy consumption viewed (although not in great detail) giving the ability to control via Apple’s own Home app or via the WIFIPLUG app. The one thing that I hadn’t spotted first time around is that it’s preferred to add the unit to the WIFIPLUG app (which then automatically adds to Home) whereas if you add to Home directly then you lose the ability to connect to the WIFIPLUG app later.

I was sufficiently impressed to buy a second unit even before I’d integrated the first, and indeed I now have another two on order.