def diff_time_span(is_train=True):
"""
不同时间粒度的预测能力
:param is_train:
:return:
"""
time_span = {
'day': (1, 12 + 31),
'week': (1, 12),
# 'month': (1, 12)
}
conf = lstm.get_init_config()
conf['model_conf']['lstm_conf']['layers'] = [300] * 3
conf['model_conf']['dense_conf']['layers'] = [1024]
conf['model_conf']['batch_size'] = 12
for ts in time_span:
conf['model_conf']['time_span'] = ts
conf['model_conf']['dense_conf']['date_features_shape'] = time_span[ts]
conf['model_conf']['model_path'] = conf['model_conf']['model_root_path'] + '/DifferentTimeSpan/model_time_span_' + ts + '.best.json'
conf['model_conf']['weight_path'] = conf['model_conf']['model_root_path'] + '/DifferentTimeSpan/weights_time_span_' + ts + '.best.hdf5'
if is_train:
lstm.train(conf)
else:
conf['model_conf']['test_split'] = 1.0
lstm.predict(conf, save_result=True, show_fitting_curve=True)
def fixed_time_lags(is_train=True):
"""
对比固定时滞（窗口）大小提取特征
:param is_train:
:return:
"""
fixed_size = 5
fixed_time_steps = {
'PM25': fixed_size,
'Press': fixed_size,
'RH': fixed_size,
'Temp': fixed_size,
'Wind Speed': fixed_size,
'Wind Direction': fixed_size,
'CO': fixed_size,
'NO2': fixed_size,
'O3': fixed_size,
'SO2': fixed_size,
'PM10': fixed_size
}
conf = lstm.get_init_config()
conf['model_conf']['lstm_conf']['layers'] = [300] * 3
conf['model_conf']['dense_conf']['layers'] = [1024]
conf['model_conf']['time_steps'] = fixed_time_steps
conf['model_conf']['lstm_conf']['max_feature_length'] = fixed_size * len(conf['model_conf']['time_steps'])
conf['model_conf']['lstm_conf']['input_shape'] = (1, fixed_size * len(conf['model_conf']['time_steps']))
conf['model_conf']['model_path'] = conf['model_conf']['model_root_path'] + '/FixedTimeLags/model_fixed_time_lags_.best.json'
conf['model_conf']['weight_path'] = conf['model_conf']['model_root_path'] + '/FixedTimeLags/weights_fixed_time_lags_.best.hdf5'
if is_train:
lstm.train(conf)
else:
lstm.predict(conf, save_result=True)
def effect_of_other_series_feature(is_train=True):
"""
其他时序特征对预测结果的影响
:param is_train:
:return:
"""
features = {
'pm25': {
'cols': ['PM25', 'Month', 'Day', 'Hour'],
'time_steps': {'PM25': 5}
},
'pm25_meteorology': {
'cols': ['PM25', 'Press', 'RH', 'Temp', 'Wind Speed', 'Wind Direction', 'Month', 'Day', 'Hour'],
'time_steps': {
'PM25': 5,
'Press': 24,
'RH': 8,
'Temp': 24,
'Wind Speed': 11,
'Wind Direction': 11
}
},
'pm25_other_pollutant': {
'cols': ['PM25', 'CO', 'NO2', 'O3', 'SO2', 'PM10', 'Month', 'Day', 'Hour'],
'time_steps': {
'PM25': 5,
'CO': 10,
'NO2': 8,
'O3': 7,
'SO2': 4,
'PM10': 4
}
}
}
conf = lstm.get_init_config()
conf['model_conf']['lstm_conf']['layers'] = [300] * 3
conf['model_conf']['dense_conf']['layers'] = [1024]
for f in features:
conf['data_conf']['use_cols'] = features[f]['cols']
conf['model_conf']['time_steps'] = features[f]['time_steps']
conf['model_conf']['lstm_conf']['max_feature_length'] = 24 * len(conf['model_conf']['time_steps'])
conf['model_conf']['lstm_conf']['input_shape'] = (1, 24 * len(conf['model_conf']['time_steps']))
conf['model_conf']['model_path'] = conf['model_conf']['model_root_path'] + '/EffectOfOtherSeriesFeature/model_' + f + '_.best.json'
conf['model_conf']['weight_path'] = conf['model_conf']['model_root_path'] + '/EffectOfOtherSeriesFeature/weights_' + f + '_.best.hdf5'
if is_train:
lstm.train(conf)
else:
conf['model_conf']['test_split'] = 1.0
lstm.predict(conf, save_result=True)
def find_best_dense_layers_nodes(is_train=True):
"""
寻找全链接层的最优层数和每层节点数
:param is_train:
:return:
"""
layers = [1, 2]
nodes = [128, 256, 512]
conf = lstm.get_init_config()
conf['model_conf']['lstm_conf']['layers'] = [] * 3
for l in layers:
for n in nodes:
conf['model_conf']['dense_conf']['layers'] = [n] * l
conf['model_conf']['model_path'] = conf['model_conf']['model_root_path'] + '/FindBestDense/model_dense_layers' + str(l) + '_nodes' + str(n) + '.json'
conf['model_conf']['weight_path'] = conf['model_conf']['model_root_path'] + '/FindBestDense/weights_dense_layers' + str(l) + '_nodes' + str(n) + '.best.hdf5'
if is_train:
lstm.train(conf)
else:
conf['model_conf']['test_split'] = 1.0
lstm.predict(conf, save_result=True)
def find_best_lstm_layers_nodes(is_train=True):
"""
寻找lstm块的最优层数和每层节点数
:param is_train:
:return:
"""
# layers = [2, 3, 4]
# nodes = [50, 100, 200, 300, 500]
layers = [5, 6, 7]
nodes = [50]
conf = lstm.get_init_config()
for l in layers:
for n in nodes:
conf['model_conf']['lstm_conf']['layers'] = [n] * l
conf['model_conf']['model_path'] = conf['model_conf']['model_root_path'] + '/FindBestLSTMBlock/model_lstm_block_layers' + str(l) + '_nodes' + str(n) + '.json'
conf['model_conf']['weight_path'] = conf['model_conf']['model_root_path'] + '/FindBestLSTMBlock/weights_lstm_block_layers' + str(l) + '_nodes' + str(n) + '.best.hdf5'
if is_train:
lstm.train(conf)
else:
# conf['model_conf']['test_split'] = 1
lstm.predict(conf, save_result=True, show_fitting_curve=True)
def find_best_epoch_and_batch_size(is_train=True):
"""
寻找最优的epoch和batch_size
:param is_train:
:return:
"""
epochs = [500, 1000, 1500]
batch_sizes = [24, 48, 72, 96]
conf = lstm.get_init_config()
conf['model_conf']['lstm_conf']['layers'] = [] * 3
conf['model_conf']['dense_conf']['layers'] = []
for e in epochs:
for bs in batch_sizes:
conf['model_conf']['epochs'] = e
conf['model_conf']['batch_size'] = bs
conf['model_conf']['model_path'] = conf['model_conf']['model_root_path'] + '/FindBestDense/model_epochs' + str(e) + '_batch' + str(bs) + '.json'
conf['model_conf']['weight_path'] = conf['model_conf']['model_root_path'] + '/FindBestDense/weights_epochs' + str(e) + '_batch' + str(bs) + '.best.hdf5'
if is_train:
lstm.train(conf)
else:
conf['model_conf']['test_split'] = 1.0
lstm.predict(conf, save_result=True)
def train_in_other_sites(is_train=True):
"""
模型在其他站点数据上训练
:param is_train:
:return:
"""
time_steps = {
'PM25': 5,
'Press': 24,
'RH': 8,
'Temp': 24,
'Wind Speed': 11,
'Wind Direction': 11,
'CO': 10,
'NO2': 8,
'O3': 7,
'SO2': 4,
'PM10': 4,
}
# stations = ['060376012', '060371103', '060374004']
stations = ['261630001']
conf = lstm.get_init_config()
conf['model_conf']['lstm_conf']['layers'] = [300] * 3
conf['model_conf']['dense_conf']['layers'] = [1024]
conf['model_conf']['is_drop_outlier'] = True
# conf['model_conf']['time_steps'] = time_steps
# conf['model_conf']['lstm_conf']['input_shape'] = (1, np.sum(list(time_steps.values())))
for s in stations:
conf['data_conf']['data_path'] = '../DataSet/Processed/Train/' + s + '_2016_2017_v1.csv'
conf['model_conf']['epochs'] = 500
conf['model_conf']['model_path'] = conf['model_conf']['model_root_path'] + '/TrainInOtherSite/model_in_' + s + '.json'
conf['model_conf']['weight_path'] = conf['model_conf']['model_root_path'] + '/TrainInOtherSite/weights_in_' + s + '.best.hdf5'
if is_train:
conf['model_conf']['test_split'] = 0.3
lstm.train(conf)
else:
conf['model_conf']['test_split'] = 1
lstm.predict(conf, save_result=True, show_fitting_curve=True)
def different_future_time_point_predict():
"""
用过去的数据预测未来不同时间点的值
:return:
"""
print('用过去的数据预测未来不同时间点的值....')
times = [0, 1, 2, 3, 4, 6, 8, 10, 12, 14, 18, 22, 23, 24, 25, 26, 27, 28, 29, 30]
conf = lstm.get_init_config()
conf['model_conf']['test_split'] = 1.0
conf['model_conf']['model_path'] = conf['model_conf']['model_root_path'] + '/model.best.json'
conf['model_conf']['weight_path'] = conf['model_conf']['model_root_path'] + '/weights.best.hdf5'
metrics = pd.DataFrame()
for t in times:
conf['model_conf']['target_offset'] = t
result = lstm.predict(conf)
df_temp = result['all_metrics']
df_temp['Future Time'] = t
metrics = metrics.append(df_temp)
metrics.to_csv(conf['model_conf']['model_root_path'] + '/DifferentFutureTime/use_best_model_predict_diff_future_time_point.csv', index=False)
def use_best_model_predict_other_sites():
"""
用最好的模型预测其他站点，查看模型对其他站点的泛化性能
:return:
"""
stations = ['261630001', '060376012', '060371103', '060374004']
years = ['2016', '2017']
conf = lstm.get_init_config()
conf['model_conf']['test_split'] = 1.0
conf['model_conf']['model_path'] = conf['model_conf']['model_root_path'] + '/model.best.json'
conf['model_conf']['weight_path'] = conf['model_conf']['model_root_path'] + '/weights.best.hdf5'
metrics = pd.DataFrame()
for s in stations:
for y in years:
conf['data_conf']['data_path'] = '../DataSet/Processed/Train/' + s + '_' + y + '_v1.csv'
print(conf['data_conf']['data_path'])
result = lstm.predict(conf)
df_temp = result['all_metrics']
df_temp['Station'] = s
df_temp['Year'] = y
metrics = metrics.append(df_temp)
metrics.to_csv(conf['model_conf']['model_root_path'] + '/PredictOtherSites/use_best_model_predict_other_sites.csv', index=False)