Пример #1
0
def run_operate(model_data, timings, backend, build_only):
    """
    For use when mode is 'operate', to allow the model to be built, edited, and
    iteratively run within Pyomo.

    """
    log_time(logger,
             timings,
             'run_start',
             comment='Backend: starting model run in operational mode')

    defaults = AttrDict.from_yaml_string(model_data.attrs['defaults'])
    run_config = AttrDict.from_yaml_string(model_data.attrs['run_config'])

    operate_params = ['purchased'] + [
        i.replace('_max', '') for i in defaults if i[-4:] == '_max'
    ]

    # Capacity results (from plan mode) can be used as the input to operate mode
    if (any(model_data.filter_by_attrs(is_result=1).data_vars)
            and run_config.get('operation.use_cap_results', False)):
        # Anything with is_result = 1 will be ignored in the Pyomo model
        for varname, varvals in model_data.data_vars.items():
            if varname in operate_params:
                varvals.attrs['is_result'] = 1
                varvals.attrs['operate_param'] = 1

    else:
        cap_max = xr.merge([
            v.rename(k.replace('_max', ''))
            for k, v in model_data.data_vars.items() if '_max' in k
        ])
        cap_equals = xr.merge([
            v.rename(k.replace('_equals', ''))
            for k, v in model_data.data_vars.items() if '_equals' in k
        ])
        caps = cap_max.update(cap_equals)
        for cap in caps.data_vars.values():
            cap.attrs['is_result'] = 1
            cap.attrs['operate_param'] = 1
        model_data.update(caps)

    # Storage initial is carried over between iterations, so must be defined along with storage
    if ('loc_techs_store' in model_data.dims.keys()
            and 'storage_initial' not in model_data.data_vars.keys()):
        model_data['storage_initial'] = (xr.DataArray(
            [0 for loc_tech in model_data.loc_techs_store.values],
            dims='loc_techs_store'))
        model_data['storage_initial'].attrs['is_result'] = 0
        exceptions.warn(
            'Initial stored energy not defined, set to zero for all '
            'loc::techs in loc_techs_store, for use in iterative optimisation')
    # Operated units is carried over between iterations, so must be defined in a milp model
    if ('loc_techs_milp' in model_data.dims.keys()
            and 'operated_units' not in model_data.data_vars.keys()):
        model_data['operated_units'] = (xr.DataArray(
            [0 for loc_tech in model_data.loc_techs_milp.values],
            dims='loc_techs_milp'))
        model_data['operated_units'].attrs['is_result'] = 1
        model_data['operated_units'].attrs['operate_param'] = 1
        exceptions.warn(
            'daily operated units not defined, set to zero for all '
            'loc::techs in loc_techs_milp, for use in iterative optimisation')

    comments, warnings, errors = checks.check_operate_params(model_data)
    exceptions.print_warnings_and_raise_errors(warnings=warnings,
                                               errors=errors)

    # Initialize our variables
    solver = run_config['solver']
    solver_io = run_config.get('solver_io', None)
    solver_options = run_config.get('solver_options', None)
    save_logs = run_config.get('save_logs', None)
    window = run_config['operation']['window']
    horizon = run_config['operation']['horizon']
    window_to_horizon = horizon - window

    # get the cumulative sum of timestep resolution, to find where we hit our window and horizon
    timestep_cumsum = model_data.timestep_resolution.cumsum(
        'timesteps').to_pandas()
    # get the timesteps at which we start and end our windows
    window_ends = timestep_cumsum.where((timestep_cumsum % window == 0) | (
        timestep_cumsum == timestep_cumsum[-1]))
    window_starts = timestep_cumsum.where((~np.isnan(window_ends.shift(1))) | (
        timestep_cumsum == timestep_cumsum[0])).dropna()

    window_ends = window_ends.dropna()
    horizon_ends = timestep_cumsum[timestep_cumsum.isin(window_ends.values +
                                                        window_to_horizon)]

    if not any(window_starts):
        raise exceptions.ModelError(
            'Not enough timesteps or incorrect timestep resolution to run in '
            'operational mode with an optimisation window of {}'.format(
                window))

    # We will only update timseries parameters
    timeseries_data_vars = [
        k for k, v in model_data.data_vars.items()
        if 'timesteps' in v.dims and v.attrs['is_result'] == 0
    ]

    # Loop through each window, solve over the horizon length, and add result to
    # result_array we only go as far as the end of the last horizon, which may
    # clip the last bit of data
    result_array = []
    # track whether each iteration finds an optimal solution or not
    terminations = []

    if build_only:
        iterations = [0]
    else:
        iterations = range(len(window_starts))

    for i in iterations:
        start_timestep = window_starts.index[i]

        # Build full model in first instance
        if i == 0:
            warmstart = False
            end_timestep = horizon_ends.index[i]
            timesteps = slice(start_timestep, end_timestep)
            window_model_data = model_data.loc[dict(timesteps=timesteps)]

            log_time(logger,
                     timings,
                     'model_gen_1',
                     comment='Backend: generating initial model')

            backend_model = backend.generate_model(window_model_data)

        # Build the full model in the last instance(s),
        # where number of timesteps is less than the horizon length
        elif i > len(horizon_ends) - 1:
            warmstart = False
            end_timestep = window_ends.index[i]
            timesteps = slice(start_timestep, end_timestep)
            window_model_data = model_data.loc[dict(timesteps=timesteps)]

            log_time(
                logger,
                timings,
                'model_gen_{}'.format(i + 1),
                comment=(
                    'Backend: iteration {}: generating new model for '
                    'end of timeseries, with horizon = {} timesteps'.format(
                        i + 1, window_ends[i] - window_starts[i])))

            backend_model = backend.generate_model(window_model_data)

        # Update relevent Pyomo Params in intermediate instances
        else:
            warmstart = True
            end_timestep = horizon_ends.index[i]
            timesteps = slice(start_timestep, end_timestep)
            window_model_data = model_data.loc[dict(timesteps=timesteps)]

            log_time(
                logger,
                timings,
                'model_gen_{}'.format(i + 1),
                comment='Backend: iteration {}: updating model parameters'.
                format(i + 1))
            # Pyomo model sees the same timestamps each time, we just change the
            # values associated with those timestamps
            for var in timeseries_data_vars:
                # New values
                var_series = window_model_data[var].to_series().dropna(
                ).replace('inf', np.inf)
                # Same timestamps
                var_series.index = backend_model.__calliope_model_data['data'][
                    var].keys()
                var_dict = var_series.to_dict()
                # Update pyomo Param with new dictionary

                getattr(backend_model, var).store_values(var_dict)

        if not build_only:
            log_time(logger,
                     timings,
                     'model_run_{}'.format(i + 1),
                     time_since_run_start=True,
                     comment='Backend: iteration {}: sending model to solver'.
                     format(i + 1))
            # After iteration 1, warmstart = True, which should speed up the process
            # Note: Warmstart isn't possible with GLPK (dealt with later on)
            _results = backend.solve_model(
                backend_model,
                solver=solver,
                solver_io=solver_io,
                solver_options=solver_options,
                save_logs=save_logs,
                warmstart=warmstart,
            )

            log_time(logger,
                     timings,
                     'run_solver_exit_{}'.format(i + 1),
                     time_since_run_start=True,
                     comment='Backend: iteration {}: solver finished running'.
                     format(i + 1))
            # xarray dataset is built for each iteration
            _termination = backend.load_results(backend_model, _results)
            terminations.append(_termination)

            _results = backend.get_result_array(backend_model, model_data)

            # We give back the actual timesteps for this iteration and take a slice
            # equal to the window length
            _results['timesteps'] = window_model_data.timesteps.copy()

            # We always save the window data. Until the last window(s) this will crop
            # the window_to_horizon timesteps. In the last window(s), optimistion will
            # only be occurring over a window length anyway
            _results = _results.loc[dict(
                timesteps=slice(None, window_ends.index[i]))]
            result_array.append(_results)

            # Set up initial storage for the next iteration
            if 'loc_techs_store' in model_data.dims.keys():
                storage_initial = _results.storage.loc[{
                    'timesteps':
                    window_ends.index[i]
                }].drop('timesteps')
                model_data['storage_initial'].loc[
                    storage_initial.coords] = storage_initial.values
                backend_model.storage_initial.store_values(
                    storage_initial.to_series().dropna().to_dict())

            # Set up total operated units for the next iteration
            if 'loc_techs_milp' in model_data.dims.keys():
                operated_units = _results.operating_units.sum(
                    'timesteps').astype(np.int)
                model_data['operated_units'].loc[{}] += operated_units.values
                backend_model.operated_units.store_values(
                    operated_units.to_series().dropna().to_dict())

            log_time(logger,
                     timings,
                     'run_solver_exit_{}'.format(i + 1),
                     time_since_run_start=True,
                     comment='Backend: iteration {}: generated solution array'.
                     format(i + 1))

    if build_only:
        results = xr.Dataset()
    else:
        # Concatenate results over the timestep dimension to get a single
        # xarray Dataset of interest
        results = xr.concat(result_array, dim='timesteps')
        if all(i == 'optimal' for i in terminations):
            results.attrs['termination_condition'] = 'optimal'
        elif all(i in ['optimal', 'feasible'] for i in terminations):
            results.attrs['termination_condition'] = 'feasible'
        else:
            results.attrs['termination_condition'] = ','.join(terminations)

        log_time(logger,
                 timings,
                 'run_solution_returned',
                 time_since_run_start=True,
                 comment='Backend: generated full solution array')

    return results, backend_model
Пример #2
0
def run_operate(model_data, timings, backend, build_only):
    """
    For use when mode is 'operate', to allow the model to be built, edited, and
    iteratively run within Pyomo.

    """
    log_time(
        logger,
        timings,
        "run_start",
        comment="Backend: starting model run in operational mode",
    )

    defaults = UpdateObserverDict(
        initial_yaml_string=model_data.attrs["defaults"],
        name="defaults",
        observer=model_data,
    )
    run_config = UpdateObserverDict(
        initial_yaml_string=model_data.attrs["run_config"],
        name="run_config",
        observer=model_data,
    )

    # New param defaults = old maximum param defaults (e.g. energy_cap gets default from energy_cap_max)
    operate_params = {
        k.replace("_max", ""): v
        for k, v in defaults.items() if k.endswith("_max")
    }
    operate_params[
        "purchased"] = 0  # no _max to work from here, so we hardcode a default

    defaults.update(operate_params)

    # Capacity results (from plan mode) can be used as the input to operate mode
    if any(model_data.filter_by_attrs(
            is_result=1).data_vars) and run_config.get(
                "operation.use_cap_results", False):
        # Anything with is_result = 1 will be ignored in the Pyomo model
        for varname, varvals in model_data.data_vars.items():
            if varname in operate_params.keys():
                varvals.attrs["is_result"] = 1
                varvals.attrs["operate_param"] = 1

    else:
        cap_max = xr.merge([
            v.rename(k.replace("_max", ""))
            for k, v in model_data.data_vars.items() if "_max" in k
        ])
        cap_equals = xr.merge([
            v.rename(k.replace("_equals", ""))
            for k, v in model_data.data_vars.items() if "_equals" in k
        ])
        caps = cap_max.update(cap_equals)
        for cap in caps.data_vars.values():
            cap.attrs["is_result"] = 1
            cap.attrs["operate_param"] = 1
        model_data.update(caps)

    comments, warnings, errors = checks.check_operate_params(model_data)
    exceptions.print_warnings_and_raise_errors(warnings=warnings,
                                               errors=errors)

    # Initialize our variables
    solver = run_config["solver"]
    solver_io = run_config.get("solver_io", None)
    solver_options = run_config.get("solver_options", None)
    save_logs = run_config.get("save_logs", None)
    window = run_config["operation"]["window"]
    horizon = run_config["operation"]["horizon"]
    window_to_horizon = horizon - window

    # get the cumulative sum of timestep resolution, to find where we hit our window and horizon
    timestep_cumsum = model_data.timestep_resolution.cumsum(
        "timesteps").to_pandas()
    # get the timesteps at which we start and end our windows
    window_ends = timestep_cumsum.where((timestep_cumsum % window == 0) | (
        timestep_cumsum == timestep_cumsum[-1]))
    window_starts = timestep_cumsum.where((~np.isnan(window_ends.shift(1))) | (
        timestep_cumsum == timestep_cumsum[0])).dropna()

    window_ends = window_ends.dropna()
    horizon_ends = timestep_cumsum[timestep_cumsum.isin(window_ends.values +
                                                        window_to_horizon)]

    if not any(window_starts):
        raise exceptions.ModelError(
            "Not enough timesteps or incorrect timestep resolution to run in "
            "operational mode with an optimisation window of {}".format(
                window))

    # We will only update timseries parameters
    timeseries_data_vars = [
        k for k, v in model_data.data_vars.items()
        if "timesteps" in v.dims and v.attrs["is_result"] == 0
    ]

    # Loop through each window, solve over the horizon length, and add result to
    # result_array we only go as far as the end of the last horizon, which may
    # clip the last bit of data
    result_array = []
    # track whether each iteration finds an optimal solution or not
    terminations = []

    if build_only:
        iterations = [0]
    else:
        iterations = range(len(window_starts))

    for i in iterations:
        start_timestep = window_starts.index[i]

        # Build full model in first instance
        if i == 0:
            warmstart = False
            end_timestep = horizon_ends.index[i]
            timesteps = slice(start_timestep, end_timestep)
            window_model_data = model_data.loc[dict(timesteps=timesteps)]

            log_time(
                logger,
                timings,
                "model_gen_1",
                comment="Backend: generating initial model",
            )

            backend_model = backend.generate_model(window_model_data)

        # Build the full model in the last instance(s),
        # where number of timesteps is less than the horizon length
        elif i > len(horizon_ends) - 1:
            warmstart = False
            end_timestep = window_ends.index[i]
            timesteps = slice(start_timestep, end_timestep)
            window_model_data = model_data.loc[dict(timesteps=timesteps)]

            log_time(
                logger,
                timings,
                "model_gen_{}".format(i + 1),
                comment=(
                    "Backend: iteration {}: generating new model for "
                    "end of timeseries, with horizon = {} timesteps".format(
                        i + 1, window_ends[i] - window_starts[i])),
            )

            backend_model = backend.generate_model(window_model_data)

        # Update relevent Pyomo Params in intermediate instances
        else:
            warmstart = True
            end_timestep = horizon_ends.index[i]
            timesteps = slice(start_timestep, end_timestep)
            window_model_data = model_data.loc[dict(timesteps=timesteps)]

            log_time(
                logger,
                timings,
                "model_gen_{}".format(i + 1),
                comment="Backend: iteration {}: updating model parameters".
                format(i + 1),
            )
            # Pyomo model sees the same timestamps each time, we just change the
            # values associated with those timestamps
            for var in timeseries_data_vars:
                # New values
                var_series = (
                    window_model_data[var].to_series().dropna().replace(
                        "inf", np.inf))
                # Same timestamps
                var_series.index = backend_model.__calliope_model_data["data"][
                    var].keys()
                var_dict = var_series.to_dict()
                # Update pyomo Param with new dictionary

                getattr(backend_model, var).store_values(var_dict)

        if not build_only:
            log_time(
                logger,
                timings,
                "model_run_{}".format(i + 1),
                time_since_run_start=True,
                comment="Backend: iteration {}: sending model to solver".
                format(i + 1),
            )
            # After iteration 1, warmstart = True, which should speed up the process
            # Note: Warmstart isn't possible with GLPK (dealt with later on)
            _results = backend.solve_model(
                backend_model,
                solver=solver,
                solver_io=solver_io,
                solver_options=solver_options,
                save_logs=save_logs,
                warmstart=warmstart,
            )

            log_time(
                logger,
                timings,
                "run_solver_exit_{}".format(i + 1),
                time_since_run_start=True,
                comment="Backend: iteration {}: solver finished running".
                format(i + 1),
            )
            # xarray dataset is built for each iteration
            _termination = backend.load_results(backend_model, _results)
            terminations.append(_termination)

            _results = backend.get_result_array(backend_model, model_data)

            # We give back the actual timesteps for this iteration and take a slice
            # equal to the window length
            _results["timesteps"] = window_model_data.timesteps.copy()

            # We always save the window data. Until the last window(s) this will crop
            # the window_to_horizon timesteps. In the last window(s), optimistion will
            # only be occurring over a window length anyway
            _results = _results.loc[dict(
                timesteps=slice(None, window_ends.index[i]))]
            result_array.append(_results)

            # Set up initial storage for the next iteration
            if "loc_techs_store" in model_data.dims.keys():
                storage_initial = _results.storage.loc[{
                    "timesteps":
                    window_ends.index[i]
                }].drop("timesteps")
                model_data["storage_initial"].loc[
                    storage_initial.coords] = storage_initial.values
                backend_model.storage_initial.store_values(
                    storage_initial.to_series().dropna().to_dict())

            # Set up total operated units for the next iteration
            if "loc_techs_milp" in model_data.dims.keys():
                operated_units = _results.operating_units.sum(
                    "timesteps").astype(np.int)
                model_data["operated_units"].loc[{}] += operated_units.values
                backend_model.operated_units.store_values(
                    operated_units.to_series().dropna().to_dict())

            log_time(
                logger,
                timings,
                "run_solver_exit_{}".format(i + 1),
                time_since_run_start=True,
                comment="Backend: iteration {}: generated solution array".
                format(i + 1),
            )

    if build_only:
        results = xr.Dataset()
    else:
        # Concatenate results over the timestep dimension to get a single
        # xarray Dataset of interest
        results = xr.concat(result_array, dim="timesteps")
        if all(i == "optimal" for i in terminations):
            results.attrs["termination_condition"] = "optimal"
        elif all(i in ["optimal", "feasible"] for i in terminations):
            results.attrs["termination_condition"] = "feasible"
        else:
            results.attrs["termination_condition"] = ",".join(terminations)

        log_time(
            logger,
            timings,
            "run_solution_returned",
            time_since_run_start=True,
            comment="Backend: generated full solution array",
        )

    return results, backend_model
Пример #3
0
def run_operate(model_data, timings, backend, build_only):
    """
    For use when mode is 'operate', to allow the model to be built, edited, and
    iteratively run within Pyomo.

    """
    log_time(timings, 'run_start',
             comment='Backend: starting model run in operational mode')

    defaults = ruamel.yaml.load(model_data.attrs['defaults'], Loader=ruamel.yaml.Loader)
    operate_params = ['purchased'] + [
        i.replace('_max', '') for i in defaults if i[-4:] == '_max'
    ]

    # Capacity results (from plan mode) can be used as the input to operate mode
    if (any(model_data.filter_by_attrs(is_result=1).data_vars) and
            model_data.attrs.get('run.operation.use_cap_results', False)):
        # Anything with is_result = 1 will be ignored in the Pyomo model
        for varname, varvals in model_data.data_vars.items():
            if varname in operate_params:
                varvals.attrs['is_result'] = 1
                varvals.attrs['operate_param'] = 1

    else:
        cap_max = xr.merge([
            v.rename(k.replace('_max', ''))
            for k, v in model_data.data_vars.items() if '_max' in k
        ])
        cap_equals = xr.merge([
            v.rename(k.replace('_equals', ''))
            for k, v in model_data.data_vars.items() if '_equals' in k
        ])
        caps = cap_max.update(cap_equals)
        for cap in caps.data_vars.values():
            cap.attrs['is_result'] = 1
            cap.attrs['operate_param'] = 1
        model_data.update(caps)

    # Storage initial is carried over between iterations, so must be defined along with storage
    if ('loc_techs_store' in model_data.dims.keys() and
        'storage_initial' not in model_data.data_vars.keys()):
        model_data['storage_initial'] = (
            xr.DataArray([0 for loc_tech in model_data.loc_techs_store.values],
                         dims='loc_techs_store')
        )
        model_data['storage_initial'].attrs['is_result'] = 0
        exceptions.ModelWarning(
            'Initial stored energy not defined, set to zero for all '
            'loc::techs in loc_techs_store, for use in iterative optimisation'
        )
    # Operated units is carried over between iterations, so must be defined in a milp model
    if ('loc_techs_milp' in model_data.dims.keys() and
        'operated_units' not in model_data.data_vars.keys()):
        model_data['operated_units'] = (
            xr.DataArray([0 for loc_tech in model_data.loc_techs_milp.values],
                         dims='loc_techs_milp')
        )
        model_data['operated_units'].attrs['is_result'] = 1
        model_data['operated_units'].attrs['operate_param'] = 1
        exceptions.ModelWarning(
            'daily operated units not defined, set to zero for all '
            'loc::techs in loc_techs_milp, for use in iterative optimisation'
        )

    comments, warnings, errors = checks.check_operate_params(model_data)
    exceptions.print_warnings_and_raise_errors(warnings=warnings, errors=errors)

    # Initialize our variables
    solver = model_data.attrs['run.solver']
    solver_io = model_data.attrs.get('run.solver_io', None)
    solver_options = model_data.attrs.get('run.solver_options', None)
    save_logs = model_data.attrs.get('run.save_logs', None)
    window = model_data.attrs['run.operation.window']
    horizon = model_data.attrs['run.operation.horizon']
    window_to_horizon = horizon - window

    # get the cumulative sum of timestep resolution, to find where we hit our window and horizon
    timestep_cumsum = model_data.timestep_resolution.cumsum('timesteps').to_pandas()
    # get the timesteps at which we start and end our windows
    window_ends = timestep_cumsum.where(
        (timestep_cumsum % window == 0) | (timestep_cumsum == timestep_cumsum[-1])
    )
    window_starts = timestep_cumsum.where(
        (~np.isnan(window_ends.shift(1))) | (timestep_cumsum == timestep_cumsum[0])
    ).dropna()

    window_ends = window_ends.dropna()
    horizon_ends = timestep_cumsum[timestep_cumsum.isin(window_ends.values + window_to_horizon)]

    if not any(window_starts):
        raise exceptions.ModelError(
            'Not enough timesteps or incorrect timestep resolution to run in '
            'operational mode with an optimisation window of {}'.format(window)
        )

    # We will only update timseries parameters
    timeseries_data_vars = [
        k for k, v in model_data.data_vars.items() if 'timesteps' in v.dims
        and v.attrs['is_result'] == 0
    ]

    # Loop through each window, solve over the horizon length, and add result to
    # result_array we only go as far as the end of the last horizon, which may
    # clip the last bit of data
    result_array = []
    # track whether each iteration finds an optimal solution or not
    terminations = []

    if build_only:
        iterations = [0]
    else:
        iterations = range(len(window_starts))

    for i in iterations:
        start_timestep = window_starts.index[i]

        # Build full model in first instance
        if i == 0:
            warmstart = False
            end_timestep = horizon_ends.index[i]
            timesteps = slice(start_timestep, end_timestep)
            window_model_data = model_data.loc[dict(timesteps=timesteps)]

            log_time(
                timings, 'model_gen_1',
                comment='Backend: generating initial model'
            )

            backend_model = backend.generate_model(window_model_data)

        # Build the full model in the last instance(s),
        # where number of timesteps is less than the horizon length
        elif i > len(horizon_ends) - 1:
            warmstart = False
            end_timestep = window_ends.index[i]
            timesteps = slice(start_timestep, end_timestep)
            window_model_data = model_data.loc[dict(timesteps=timesteps)]

            log_time(
                timings, 'model_gen_{}'.format(i + 1),
                comment=(
                    'Backend: iteration {}: generating new model for '
                    'end of timeseries, with horizon = {} timesteps'
                    .format(i + 1, window_ends[i] - window_starts[i])
                )
            )

            backend_model = backend.generate_model(window_model_data)

        # Update relevent Pyomo Params in intermediate instances
        else:
            warmstart = True
            end_timestep = horizon_ends.index[i]
            timesteps = slice(start_timestep, end_timestep)
            window_model_data = model_data.loc[dict(timesteps=timesteps)]

            log_time(
                timings, 'model_gen_{}'.format(i + 1),
                comment='Backend: iteration {}: updating model parameters'.format(i + 1)
            )
            # Pyomo model sees the same timestamps each time, we just change the
            # values associated with those timestamps
            for var in timeseries_data_vars:
                # New values
                var_series = window_model_data[var].to_series().dropna().replace('inf', np.inf)
                # Same timestamps
                var_series.index = backend_model.__calliope_model_data__['data'][var].keys()
                var_dict = var_series.to_dict()
                # Update pyomo Param with new dictionary
                for k, v in getattr(backend_model, var).items():
                    if k in var_dict:
                        v.set_value(var_dict[k])

        if not build_only:
            log_time(
                timings, 'model_run_{}'.format(i + 1), time_since_start=True,
                comment='Backend: iteration {}: sending model to solver'.format(i + 1)
            )
            # After iteration 1, warmstart = True, which should speed up the process
            # Note: Warmstart isn't possible with GLPK (dealt with later on)
            _results = backend.solve_model(
                backend_model, solver=solver, solver_io=solver_io,
                solver_options=solver_options, save_logs=save_logs, warmstart=warmstart,
            )

            log_time(
                timings, 'run_solver_exit_{}'.format(i + 1), time_since_start=True,
                comment='Backend: iteration {}: solver finished running'.format(i + 1)
            )
            # xarray dataset is built for each iteration
            _termination = backend.load_results(backend_model, _results)
            terminations.append(_termination)

            _results = backend.get_result_array(backend_model, model_data)

            # We give back the actual timesteps for this iteration and take a slice
            # equal to the window length
            _results['timesteps'] = window_model_data.timesteps.copy()

            # We always save the window data. Until the last window(s) this will crop
            # the window_to_horizon timesteps. In the last window(s), optimistion will
            # only be occurring over a window length anyway
            _results = _results.loc[dict(timesteps=slice(None, window_ends.index[i]))]
            result_array.append(_results)

            # Set up initial storage for the next iteration
            if 'loc_techs_store' in model_data.dims.keys():
                storage_initial = _results.storage.loc[{'timesteps': window_ends.index[i]}].drop('timesteps')
                model_data['storage_initial'].loc[storage_initial.coords] = storage_initial.values
                for k, v in backend_model.storage_initial.items():
                    v.set_value(storage_initial.to_series().dropna().to_dict()[k])

            # Set up total operated units for the next iteration
            if 'loc_techs_milp' in model_data.dims.keys():
                operated_units = _results.operating_units.sum('timesteps').astype(np.int)
                model_data['operated_units'].loc[{}] += operated_units.values
                for k, v in backend_model.operated_units.items():
                    v.set_value(operated_units.to_series().dropna().to_dict()[k])

            log_time(
                timings, 'run_solver_exit_{}'.format(i + 1), time_since_start=True,
                comment='Backend: iteration {}: generated solution array'.format(i + 1)
            )

    if build_only:
        results = xr.Dataset()
    else:
        # Concatenate results over the timestep dimension to get a single
        # xarray Dataset of interest
        results = xr.concat(result_array, dim='timesteps')
        if all(i == 'optimal' for i in terminations):
            results.attrs['termination_condition'] = 'optimal'
        else:
            results.attrs['termination_condition'] = ','.join(terminations)

        log_time(
            timings, 'run_solution_returned', time_since_start=True,
            comment='Backend: generated full solution array'
        )

    return results, backend_model