Ejemplo n.º 1
0
    def test_speed(self):
        """
        the purpose of this test is to figure out the
        speed characteristics of qm.
        testcases of interest is
          12 (4x3)  forecast typical arome 4 times a day, use last 3 days, 1-3 hour dt, 14 days ahead
          100 historical scenarios
          time-axis wanted is
          next 14 days plus historical scenarios 3..60 weeks ahead

        expected performance:
          the first 14 days includes sorting 10 forcasts, 100 historical pr. timestep.
           the period after should be close to 'memcpy' performance.

        """
        # Arrange the inputs

        utc = Calendar()
        n_hist_ts = 100
        n_fc_days = 14
        n_hist_days = n_fc_days + 360
        n_fc_ts = 10
        t0 = utc.time(2017, 1, 1, 0, 0, 0)

        def generate_ts(ta: TimeAxis, n_fc: int) -> TsVector:
            fx_avg = ts_point_fx.POINT_AVERAGE_VALUE
            r = TsVector()
            w = 2 * 3.14 / len(ta)
            for i in range(n_fc):
                a = np.random.ranf() * 20 - 10.0
                b = np.random.ranf() * 5.0
                v = dv([a + b * math.sin(w * i) for i in range(len(ta))])
                r.append(TimeSeries(ta, v, fx_avg))
            return r

        ta_hist = TimeAxis(t0, deltahours(1), 24 * n_hist_days)
        historical_scenario_ts = generate_ts(ta_hist, n_hist_ts)
        fc_set = TsVectorSet()
        fc_weight = dv()
        n_fc_sets = 4 * 2
        fc_every_dt = deltahours(6)  # six hours between each arome fc.
        dt_fc = deltahours(1)
        for i in range(n_fc_sets):
            t0_fc = t0 + fc_every_dt * i
            fc_set.append(
                generate_ts(TimeAxis(t0_fc, dt_fc, 24 * n_fc_days), n_fc_ts))
            fc_weight.append(float(3 + i))

        # interpolation_start= no_utctime
        # Act
        interpolated_quantiles = False
        qm_end_idx1 = 24 * (n_fc_days - 2)
        qm_end_idx2 = 24 * (n_fc_days - 1)
        n_ts = 0
        n_v = 0
        print(r"n_days\ttime_used[s]\n")
        tot_seconds = 0.0
        for h_days in range(n_fc_days + 10, n_hist_days, 30):
            ta_qm = TimeAxis(t0 + n_fc_sets * fc_every_dt, dt_fc, 24 * h_days)
            a0 = utctime_now()
            result = quantile_map_forecast(fc_set, fc_weight,
                                           historical_scenario_ts, ta_qm,
                                           ta_qm.time(qm_end_idx1),
                                           ta_qm.time(qm_end_idx2),
                                           interpolated_quantiles)
            self.assertIsNotNone(result)
            n_ts += len(result)
            n_v += len(result[0]) * len(result)
            a1 = utctime_now()
            tot_seconds += float(a1 - a0)
            print(f' {h_days}\t{float(a1-a0)}')
        print(
            f'Total of {n_ts} ts, was forecasted, number values produced {n_v/1000000} Mpts, Mb/s={8.0*n_v/1000000/tot_seconds}'
        )
Ejemplo n.º 2
0
    def test_forecast(self):
        fx_avg = ts_point_fx.POINT_AVERAGE_VALUE
        utc = Calendar()
        ta = TimeAxis(utc.time(2017, 1, 1, 0, 0, 0), deltahours(24), 4)
        historical_data = TsVector()

        forecast_sets = TsVectorSet()
        weight_sets = dv()
        num_historical_data = 56

        # Let's make three sets, one of two elements, one of three, and one of
        # four.

        forecasts_1 = TsVector()
        forecasts_2 = TsVector()
        forecasts_3 = TsVector()

        forecasts_1.append(TimeSeries(ta, dv([13.4, 15.6, 17.1, 19.1]),
                                      fx_avg))
        forecasts_1.append(TimeSeries(ta, dv([34.1, 2.40, 43.9, 10.2]),
                                      fx_avg))
        forecast_sets.append(forecasts_1)
        weight_sets.append(5.0)

        forecasts_2.append(TimeSeries(ta, dv([83.1, -42.2, 0.4, 23.4]),
                                      fx_avg))
        forecasts_2.append(TimeSeries(ta, dv([15.1, 6.500, 4.2, 2.9]), fx_avg))
        forecasts_2.append(TimeSeries(ta, dv([53.1, 87.90, 23.8, 5.6]),
                                      fx_avg))
        forecast_sets.append(forecasts_2)
        weight_sets.append(9.0)

        forecasts_3.append(
            TimeSeries(ta, dv([1.5, -1.9, -17.2, -10.0]), fx_avg))
        forecasts_3.append(
            TimeSeries(ta, dv([4.7, 18.2, 15.3000, 8.9]), fx_avg))
        forecasts_3.append(
            TimeSeries(ta, dv([-45.2, -2.3, 80.2, 71.0]), fx_avg))
        forecasts_3.append(
            TimeSeries(ta, dv([45.1, -92.0, 34.4, 65.8]), fx_avg))
        forecast_sets.append(forecasts_3)
        weight_sets.append(3.0)

        for i in range(num_historical_data):
            historical_data.append(
                TimeSeries(ta,
                           dv.from_numpy(np.random.random(ta.size()) * 50.0),
                           fx_avg))

        # need one more exposed from core here: auto historical_order = qm::quantile_index<tsa_t>(historical_data, ta);

        interpolation_start = ta.time(2)
        interpolation_end = ta.time(3)
        # Act
        result = quantile_map_forecast(forecast_sets, weight_sets,
                                       historical_data, ta,
                                       interpolation_start, interpolation_end,
                                       False)

        self.assertIsNotNone(result)
        self.assertEqual(len(result), num_historical_data)
        for ts in result:
            self.assertEqual(ts.size(), ta.size())